llmling_agent

Class info

Classes

Name	Children	Inherits
Agent llmling_agent.agent.agent Agent for AI-powered interaction with LLMling resources and tools.		MessageNode TaskManagerMixin Generic
AgentConfig llmling_agent.models.agents Configuration for a single agent in the system.		NodeConfig
AgentContext llmling_agent.agent.context Runtime context for agent execution.		NodeContext Generic
AgentPool llmling_agent.delegation.pool Pool managing message processing nodes (agents and teams).		BaseRegistry Generic
AgentsManifest llmling_agent.models.manifest Complete agent configuration manifest defining all available agents.		ConfigModel
AudioBase64Content llmling_agent.models.content Audio from base64 data.		AudioContent
AudioURLContent llmling_agent.models.content Audio from URL.		AudioContent
BaseTeam llmling_agent.delegation.base_team Base class for Team and TeamRun.	Team TeamRun	MessageNode Generic
ChatMessage llmling_agent.messaging.messages Common message format for all UI types.		Generic
ImageBase64Content llmling_agent.models.content Image from base64 data.		BaseImageContent
ImageURLContent llmling_agent.models.content Image from URL.		BaseImageContent
MessageNode llmling_agent.messaging.messagenode Base class for all message processing nodes.	Agent StructuredAgent BaseTeam	MessageEmitter Generic
PDFBase64Content llmling_agent.models.content PDF from base64 data.		BasePDFContent
PDFURLContent llmling_agent.models.content PDF from URL.		BasePDFContent
StructuredAgent llmling_agent.agent.structured Wrapper for Agent that enforces a specific result type.		MessageNode Generic
Team llmling_agent.delegation.team Group of agents that can execute together.		BaseTeam Generic
TeamRun llmling_agent.delegation.teamrun Handles team operations with monitoring.		BaseTeam Generic
Tool llmling_agent.tools.base Information about a registered tool.		object
ToolCallInfo llmling_agent.tools.tool_call_info Information about an executed tool call.		Schema
VideoURLContent llmling_agent.models.content Video from URL.		VideoContent

🛈 DocStrings

Agent configuration and creation.

Agent

Bases: MessageNode[TDeps, str], TaskManagerMixin

Agent for AI-powered interaction with LLMling resources and tools.

Generically typed with: LLMLingAgent[Type of Dependencies, Type of Result]

This agent integrates LLMling's resource system with PydanticAI's agent capabilities. It provides: - Access to resources through RuntimeConfig - Tool registration for resource operations - System prompt customization - Signals - Message history management - Database logging

Source code in src/llmling_agent/agent/agent.py

class Agent[TDeps](MessageNode[TDeps, str], TaskManagerMixin):
    """Agent for AI-powered interaction with LLMling resources and tools.

    Generically typed with: LLMLingAgent[Type of Dependencies, Type of Result]

    This agent integrates LLMling's resource system with PydanticAI's agent capabilities.
    It provides:
    - Access to resources through RuntimeConfig
    - Tool registration for resource operations
    - System prompt customization
    - Signals
    - Message history management
    - Database logging
    """

    @dataclass(frozen=True)
    class AgentReset:
        """Emitted when agent is reset."""

        agent_name: AgentName
        previous_tools: dict[str, bool]
        new_tools: dict[str, bool]
        timestamp: datetime = field(default_factory=get_now)

    # this fixes weird mypy issue
    conversation: ConversationManager
    talk: Interactions
    model_changed = Signal(object)  # Model | None
    chunk_streamed = Signal(str, str)  # (chunk, message_id)
    run_failed = Signal(str, Exception)
    agent_reset = Signal(AgentReset)

    def __init__(  # noqa: PLR0915
        # we dont use AgentKwargs here so that we can work with explicit ones in the ctor
        self,
        name: str = "llmling-agent",
        provider: AgentType = "pydantic_ai",
        *,
        model: ModelType = None,
        runtime: RuntimeConfig | Config | StrPath | None = None,
        context: AgentContext[TDeps] | None = None,
        session: SessionIdType | SessionQuery | MemoryConfig | bool | int = None,
        system_prompt: AnyPromptType | Sequence[AnyPromptType] = (),
        description: str | None = None,
        tools: Sequence[ToolType] | None = None,
        capabilities: Capabilities | None = None,
        mcp_servers: Sequence[str | MCPServerConfig] | None = None,
        resources: Sequence[Resource | PromptType | str] = (),
        retries: int = 1,
        result_retries: int | None = None,
        end_strategy: EndStrategy = "early",
        defer_model_check: bool = False,
        input_provider: InputProvider | None = None,
        parallel_init: bool = True,
        debug: bool = False,
    ):
        """Initialize agent with runtime configuration.

        Args:
            runtime: Runtime configuration providing access to resources/tools
            context: Agent context with capabilities and configuration
            provider: Agent type to use (ai: PydanticAIProvider, human: HumanProvider)
            session: Memory configuration.
                - None: Default memory config
                - False: Disable message history (max_messages=0)
                - int: Max tokens for memory
                - str/UUID: Session identifier
                - SessionQuery: Query to recover conversation
                - MemoryConfig: Complete memory configuration
            model: The default model to use (defaults to GPT-4)
            system_prompt: Static system prompts to use for this agent
            name: Name of the agent for logging
            description: Description of the Agent ("what it can do")
            tools: List of tools to register with the agent
            capabilities: Capabilities for the agent
            mcp_servers: MCP servers to connect to
            resources: Additional resources to load
            retries: Default number of retries for failed operations
            result_retries: Max retries for result validation (defaults to retries)
            end_strategy: Strategy for handling tool calls that are requested alongside
                          a final result
            defer_model_check: Whether to defer model evaluation until first run
            input_provider: Provider for human input (tool confirmation / HumanProviders)
            parallel_init: Whether to initialize resources in parallel
            debug: Whether to enable debug mode
        """
        from llmling_agent.agent import AgentContext
        from llmling_agent.agent.conversation import ConversationManager
        from llmling_agent.agent.interactions import Interactions
        from llmling_agent.agent.sys_prompts import SystemPrompts
        from llmling_agent.resource_providers.capability_provider import (
            CapabilitiesResourceProvider,
        )
        from llmling_agent_providers.base import AgentProvider

        self._infinite = False
        # save some stuff for asnyc init
        self._owns_runtime = False
        # prepare context
        ctx = context or AgentContext[TDeps].create_default(
            name,
            input_provider=input_provider,
            capabilities=capabilities,
        )
        self._context = ctx
        memory_cfg = (
            session
            if isinstance(session, MemoryConfig)
            else MemoryConfig.from_value(session)
        )
        super().__init__(
            name=name,
            context=ctx,
            description=description,
            enable_logging=memory_cfg.enable,
            mcp_servers=mcp_servers,
        )
        # Initialize runtime
        match runtime:
            case None:
                ctx.runtime = RuntimeConfig.from_config(Config())
            case Config() | str() | PathLike():
                ctx.runtime = RuntimeConfig.from_config(runtime)
            case RuntimeConfig():
                ctx.runtime = runtime

        runtime_provider = RuntimePromptProvider(ctx.runtime)
        ctx.definition.prompt_manager.providers["runtime"] = runtime_provider
        # Initialize tool manager
        all_tools = list(tools or [])
        self.tools = ToolManager(all_tools)
        self.tools.add_provider(self.mcp)
        if builtin_tools := ctx.config.get_tool_provider():
            self.tools.add_provider(builtin_tools)

        # Initialize conversation manager
        resources = list(resources)
        if ctx.config.knowledge:
            resources.extend(ctx.config.knowledge.get_resources())
        self.conversation = ConversationManager(self, memory_cfg, resources=resources)
        # Initialize provider
        match provider:
            case "pydantic_ai":
                validate_import("pydantic_ai", "pydantic_ai")
                from llmling_agent_providers.pydanticai import PydanticAIProvider

                if model and not isinstance(model, str):
                    from pydantic_ai import models

                    assert isinstance(model, models.Model)
                self._provider: AgentProvider = PydanticAIProvider(
                    model=model,
                    retries=retries,
                    end_strategy=end_strategy,
                    result_retries=result_retries,
                    defer_model_check=defer_model_check,
                    debug=debug,
                    context=ctx,
                )
            case "human":
                from llmling_agent_providers.human import HumanProvider

                self._provider = HumanProvider(name=name, debug=debug, context=ctx)
            case Callable():
                from llmling_agent_providers.callback import CallbackProvider

                self._provider = CallbackProvider(
                    provider, name=name, debug=debug, context=ctx
                )
            case "litellm":
                validate_import("litellm", "litellm")
                from llmling_agent_providers.litellm_provider import LiteLLMProvider

                self._provider = LiteLLMProvider(
                    name=name,
                    debug=debug,
                    retries=retries,
                    context=ctx,
                    model=model,
                )
            case AgentProvider():
                self._provider = provider
                self._provider.context = ctx
            case _:
                msg = f"Invalid agent type: {type}"
                raise ValueError(msg)
        self.tools.add_provider(CapabilitiesResourceProvider(ctx.capabilities))

        if ctx and ctx.definition:
            from llmling_agent.observability import registry

            registry.register_providers(ctx.definition.observability)

        # init variables
        self._debug = debug
        self._result_type: type | None = None
        self.parallel_init = parallel_init
        self.name = name
        self._background_task: asyncio.Task[Any] | None = None

        # Forward provider signals
        self._provider.chunk_streamed.connect(self.chunk_streamed)
        self._provider.model_changed.connect(self.model_changed)
        self._provider.tool_used.connect(self.tool_used)
        self._provider.model_changed.connect(self.model_changed)

        self.talk = Interactions(self)

        # Set up system prompts
        config_prompts = ctx.config.system_prompts if ctx else []
        all_prompts: list[AnyPromptType] = list(config_prompts)
        if isinstance(system_prompt, list):
            all_prompts.extend(system_prompt)
        else:
            all_prompts.append(system_prompt)
        self.sys_prompts = SystemPrompts(all_prompts, context=ctx)

    def __repr__(self) -> str:
        desc = f", {self.description!r}" if self.description else ""
        tools = f", tools={len(self.tools)}" if self.tools else ""
        return f"Agent({self.name!r}, provider={self._provider.NAME!r}{desc}{tools})"

    def __prompt__(self) -> str:
        typ = self._provider.__class__.__name__
        model = self.model_name or "default"
        parts = [f"Agent: {self.name}", f"Type: {typ}", f"Model: {model}"]
        if self.description:
            parts.append(f"Description: {self.description}")
        parts.extend([self.tools.__prompt__(), self.conversation.__prompt__()])

        return "\n".join(parts)

    async def __aenter__(self) -> Self:
        """Enter async context and set up MCP servers."""
        try:
            # Collect all coroutines that need to be run
            coros: list[Coroutine[Any, Any, Any]] = []

            # Runtime initialization if needed
            runtime_ref = self.context.runtime
            if runtime_ref and not runtime_ref._initialized:
                self._owns_runtime = True
                coros.append(runtime_ref.__aenter__())

            # Events initialization
            coros.append(super().__aenter__())

            # Get conversation init tasks directly
            coros.extend(self.conversation.get_initialization_tasks())

            # Execute coroutines either in parallel or sequentially
            if self.parallel_init and coros:
                await asyncio.gather(*coros)
            else:
                for coro in coros:
                    await coro
            if runtime_ref:
                self.tools.add_provider(RuntimeResourceProvider(runtime_ref))
            for provider in await self.context.config.get_toolsets():
                self.tools.add_provider(provider)
        except Exception as e:
            # Clean up in reverse order
            if self._owns_runtime and runtime_ref and self.context.runtime == runtime_ref:
                await runtime_ref.__aexit__(type(e), e, e.__traceback__)
            msg = "Failed to initialize agent"
            raise RuntimeError(msg) from e
        else:
            return self

    async def __aexit__(
        self,
        exc_type: type[BaseException] | None,
        exc_val: BaseException | None,
        exc_tb: TracebackType | None,
    ):
        """Exit async context."""
        await super().__aexit__(exc_type, exc_val, exc_tb)
        try:
            await self.mcp.__aexit__(exc_type, exc_val, exc_tb)
        finally:
            if self._owns_runtime and self.context.runtime:
                self.tools.remove_provider("runtime")
                await self.context.runtime.__aexit__(exc_type, exc_val, exc_tb)
            # for provider in await self.context.config.get_toolsets():
            #     self.tools.remove_provider(provider.name)

    @overload
    def __and__(
        self, other: Agent[TDeps] | StructuredAgent[TDeps, Any]
    ) -> Team[TDeps]: ...

    @overload
    def __and__(self, other: Team[TDeps]) -> Team[TDeps]: ...

    @overload
    def __and__(self, other: ProcessorCallback[Any]) -> Team[TDeps]: ...

    def __and__(self, other: MessageNode[Any, Any] | ProcessorCallback[Any]) -> Team[Any]:
        """Create agent group using | operator.

        Example:
            group = analyzer & planner & executor  # Create group of 3
            group = analyzer & existing_group  # Add to existing group
        """
        from llmling_agent.agent import StructuredAgent
        from llmling_agent.delegation.team import Team

        match other:
            case Team():
                return Team([self, *other.agents])
            case Callable():
                if callable(other):
                    if has_return_type(other, str):
                        agent_2 = Agent.from_callback(other)
                    else:
                        agent_2 = StructuredAgent.from_callback(other)
                agent_2.context.pool = self.context.pool
                return Team([self, agent_2])
            case MessageNode():
                return Team([self, other])
            case _:
                msg = f"Invalid agent type: {type(other)}"
                raise ValueError(msg)

    @overload
    def __or__(self, other: MessageNode[TDeps, Any]) -> TeamRun[TDeps, Any]: ...

    @overload
    def __or__[TOtherDeps](
        self,
        other: MessageNode[TOtherDeps, Any],
    ) -> TeamRun[Any, Any]: ...

    @overload
    def __or__(self, other: ProcessorCallback[Any]) -> TeamRun[Any, Any]: ...

    def __or__(self, other: MessageNode[Any, Any] | ProcessorCallback[Any]) -> TeamRun:
        # Create new execution with sequential mode (for piping)
        from llmling_agent import StructuredAgent, TeamRun

        if callable(other):
            if has_return_type(other, str):
                other = Agent.from_callback(other)
            else:
                other = StructuredAgent.from_callback(other)
            other.context.pool = self.context.pool

        return TeamRun([self, other])

    @classmethod
    def from_callback(
        cls,
        callback: ProcessorCallback[str],
        *,
        name: str | None = None,
        debug: bool = False,
        **kwargs: Any,
    ) -> Agent[None]:
        """Create an agent from a processing callback.

        Args:
            callback: Function to process messages. Can be:
                - sync or async
                - with or without context
                - must return str for pipeline compatibility
            name: Optional name for the agent
            debug: Whether to enable debug mode
            kwargs: Additional arguments for agent
        """
        from llmling_agent_providers.callback import CallbackProvider

        name = name or callback.__name__ or "processor"
        provider = CallbackProvider(callback, name=name)
        return Agent[None](provider=provider, name=name, debug=debug, **kwargs)

    @property
    def name(self) -> str:
        """Get agent name."""
        return self._name or "llmling-agent"

    @name.setter
    def name(self, value: str):
        self._provider.name = value
        self._name = value

    @property
    def context(self) -> AgentContext[TDeps]:
        """Get agent context."""
        return self._context

    @context.setter
    def context(self, value: AgentContext[TDeps]):
        """Set agent context and propagate to provider."""
        self._provider.context = value
        self.mcp.context = value
        self._context = value

    def set_result_type(
        self,
        result_type: type[TResult] | str | ResponseDefinition | None,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ):
        """Set or update the result type for this agent.

        Args:
            result_type: New result type, can be:
                - A Python type for validation
                - Name of a response definition
                - Response definition instance
                - None to reset to unstructured mode
            tool_name: Optional override for tool name
            tool_description: Optional override for tool description
        """
        logger.debug("Setting result type to: %s for %r", result_type, self.name)
        self._result_type = to_type(result_type)

    @property
    def provider(self) -> AgentProvider:
        """Get the underlying provider."""
        return self._provider

    @provider.setter
    def provider(self, value: AgentProvider, model: ModelType = None):
        """Set the underlying provider."""
        from llmling_agent_providers.base import AgentProvider

        name = self.name
        debug = self._debug
        self._provider.chunk_streamed.disconnect(self.chunk_streamed)
        self._provider.model_changed.disconnect(self.model_changed)
        self._provider.tool_used.disconnect(self.tool_used)
        self._provider.model_changed.disconnect(self.model_changed)
        match value:
            case AgentProvider():
                self._provider = value
            case "pydantic_ai":
                validate_import("pydantic_ai", "pydantic_ai")
                from llmling_agent_providers.pydanticai import PydanticAIProvider

                self._provider = PydanticAIProvider(model=model, name=name, debug=debug)
            case "human":
                from llmling_agent_providers.human import HumanProvider

                self._provider = HumanProvider(name=name, debug=debug)
            case "litellm":
                validate_import("litellm", "litellm")
                from llmling_agent_providers.litellm_provider import LiteLLMProvider

                self._provider = LiteLLMProvider(model=model, name=name, debug=debug)
            case Callable():
                from llmling_agent_providers.callback import CallbackProvider

                self._provider = CallbackProvider(value, name=name, debug=debug)
            case _:
                msg = f"Invalid agent type: {type}"
                raise ValueError(msg)
        self._provider.chunk_streamed.connect(self.chunk_streamed)
        self._provider.model_changed.connect(self.model_changed)
        self._provider.tool_used.connect(self.tool_used)
        self._provider.model_changed.connect(self.model_changed)
        self._provider.context = self._context

    @overload
    def to_structured(
        self,
        result_type: None,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ) -> Self: ...

    @overload
    def to_structured[TResult](
        self,
        result_type: type[TResult] | str | ResponseDefinition,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ) -> StructuredAgent[TDeps, TResult]: ...

    def to_structured[TResult](
        self,
        result_type: type[TResult] | str | ResponseDefinition | None,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ) -> StructuredAgent[TDeps, TResult] | Self:
        """Convert this agent to a structured agent.

        If result_type is None, returns self unchanged (no wrapping).
        Otherwise creates a StructuredAgent wrapper.

        Args:
            result_type: Type for structured responses. Can be:
                - A Python type (Pydantic model)
                - Name of response definition from context
                - Complete response definition
                - None to skip wrapping
            tool_name: Optional override for result tool name
            tool_description: Optional override for result tool description

        Returns:
            Either StructuredAgent wrapper or self unchanged
        from llmling_agent.agent import StructuredAgent
        """
        if result_type is None:
            return self

        from llmling_agent.agent import StructuredAgent

        return StructuredAgent(
            self,
            result_type=result_type,
            tool_name=tool_name,
            tool_description=tool_description,
        )

    def is_busy(self) -> bool:
        """Check if agent is currently processing tasks."""
        return bool(self._pending_tasks or self._background_task)

    @property
    def model_name(self) -> str | None:
        """Get the model name in a consistent format."""
        return self._provider.model_name

    def to_tool(
        self,
        *,
        name: str | None = None,
        reset_history_on_run: bool = True,
        pass_message_history: bool = False,
        share_context: bool = False,
        parent: AnyAgent[Any, Any] | None = None,
    ) -> Tool:
        """Create a tool from this agent.

        Args:
            name: Optional tool name override
            reset_history_on_run: Clear agent's history before each run
            pass_message_history: Pass parent's message history to agent
            share_context: Whether to pass parent's context/deps
            parent: Optional parent agent for history/context sharing
        """
        tool_name = f"ask_{self.name}"

        async def wrapped_tool(prompt: str) -> str:
            if pass_message_history and not parent:
                msg = "Parent agent required for message history sharing"
                raise ToolError(msg)

            if reset_history_on_run:
                self.conversation.clear()

            history = None
            if pass_message_history and parent:
                history = parent.conversation.get_history()
                old = self.conversation.get_history()
                self.conversation.set_history(history)
            result = await self.run(prompt, result_type=self._result_type)
            if history:
                self.conversation.set_history(old)
            return result.data

        normalized_name = self.name.replace("_", " ").title()
        docstring = f"Get expert answer from specialized agent: {normalized_name}"
        if self.description:
            docstring = f"{docstring}\n\n{self.description}"

        wrapped_tool.__doc__ = docstring
        wrapped_tool.__name__ = tool_name

        return Tool.from_callable(
            wrapped_tool,
            name_override=tool_name,
            description_override=docstring,
        )

    @track_action("Calling Agent.run: {prompts}:")
    async def _run(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | ChatMessage[Any],
        result_type: type[TResult] | None = None,
        model: ModelType = None,
        store_history: bool = True,
        tool_choice: str | list[str] | None = None,
        usage_limits: UsageLimits | None = None,
        message_id: str | None = None,
        conversation_id: str | None = None,
        messages: list[ChatMessage[Any]] | None = None,
        wait_for_connections: bool | None = None,
    ) -> ChatMessage[TResult]:
        """Run agent with prompt and get response.

        Args:
            prompts: User query or instruction
            result_type: Optional type for structured responses
            model: Optional model override
            store_history: Whether the message exchange should be added to the
                            context window
            tool_choice: Filter tool choice by name
            usage_limits: Optional usage limits for the model
            message_id: Optional message id for the returned message.
                        Automatically generated if not provided.
            conversation_id: Optional conversation id for the returned message.
            messages: Optional list of messages to replace the conversation history
            wait_for_connections: Whether to wait for connected agents to complete

        Returns:
            Result containing response and run information

        Raises:
            UnexpectedModelBehavior: If the model fails or behaves unexpectedly
        """
        """Run agent with prompt and get response."""
        message_id = message_id or str(uuid4())
        tools = await self.tools.get_tools(state="enabled", names=tool_choice)
        self.set_result_type(result_type)
        start_time = time.perf_counter()
        sys_prompt = await self.sys_prompts.format_system_prompt(self)

        message_history = (
            messages if messages is not None else self.conversation.get_history()
        )
        try:
            result = await self._provider.generate_response(
                *await convert_prompts(prompts),
                message_id=message_id,
                message_history=message_history,
                tools=tools,
                result_type=result_type,
                usage_limits=usage_limits,
                model=model,
                system_prompt=sys_prompt,
            )
        except Exception as e:
            logger.exception("Agent run failed")
            self.run_failed.emit("Agent run failed", e)
            raise
        else:
            response_msg = ChatMessage[TResult](
                content=result.content,
                role="assistant",
                name=self.name,
                model=result.model_name,
                message_id=message_id,
                conversation_id=conversation_id,
                tool_calls=result.tool_calls,
                cost_info=result.cost_and_usage,
                response_time=time.perf_counter() - start_time,
                provider_extra=result.provider_extra or {},
            )
            if self._debug:
                import devtools

                devtools.debug(response_msg)
            return response_msg

    @asynccontextmanager
    async def run_stream(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        result_type: type[TResult] | None = None,
        model: ModelType = None,
        tool_choice: str | list[str] | None = None,
        store_history: bool = True,
        usage_limits: UsageLimits | None = None,
        message_id: str | None = None,
        conversation_id: str | None = None,
        messages: list[ChatMessage[Any]] | None = None,
        wait_for_connections: bool | None = None,
    ) -> AsyncIterator[StreamingResponseProtocol[TResult]]:
        """Run agent with prompt and get a streaming response.

        Args:
            prompt: User query or instruction
            result_type: Optional type for structured responses
            model: Optional model override
            tool_choice: Filter tool choice by name
            store_history: Whether the message exchange should be added to the
                           context window
            usage_limits: Optional usage limits for the model
            message_id: Optional message id for the returned message.
                        Automatically generated if not provided.
            conversation_id: Optional conversation id for the returned message.
            messages: Optional list of messages to replace the conversation history
            wait_for_connections: Whether to wait for connected agents to complete

        Returns:
            A streaming result to iterate over.

        Raises:
            UnexpectedModelBehavior: If the model fails or behaves unexpectedly
        """
        message_id = message_id or str(uuid4())
        user_msg, prompts = await self.pre_run(*prompt)
        self.set_result_type(result_type)
        start_time = time.perf_counter()
        sys_prompt = await self.sys_prompts.format_system_prompt(self)
        tools = await self.tools.get_tools(state="enabled", names=tool_choice)
        message_history = (
            messages if messages is not None else self.conversation.get_history()
        )
        try:
            async with self._provider.stream_response(
                *prompts,
                message_id=message_id,
                message_history=message_history,
                result_type=result_type,
                model=model,
                store_history=store_history,
                tools=tools,
                usage_limits=usage_limits,
                system_prompt=sys_prompt,
            ) as stream:
                yield stream
                usage = stream.usage()
                cost_info = None
                model_name = stream.model_name  # type: ignore
                if model_name:
                    cost_info = await TokenCost.from_usage(
                        usage,
                        model_name,
                        str(user_msg.content),
                        str(stream.formatted_content),  # type: ignore
                    )
                response_msg = ChatMessage[TResult](
                    content=cast(TResult, stream.formatted_content),  # type: ignore
                    role="assistant",
                    name=self.name,
                    model=model_name,
                    message_id=message_id,
                    conversation_id=user_msg.conversation_id,
                    cost_info=cost_info,
                    response_time=time.perf_counter() - start_time,
                    # provider_extra=stream.provider_extra or {},
                )
                self.message_sent.emit(response_msg)
                if store_history:
                    self.conversation.add_chat_messages([user_msg, response_msg])
                await self.connections.route_message(
                    response_msg,
                    wait=wait_for_connections,
                )

        except Exception as e:
            logger.exception("Agent stream failed")
            self.run_failed.emit("Agent stream failed", e)
            raise

    async def run_iter(
        self,
        *prompt_groups: Sequence[AnyPromptType | PIL.Image.Image | os.PathLike[str]],
        result_type: type[TResult] | None = None,
        model: ModelType = None,
        store_history: bool = True,
        wait_for_connections: bool | None = None,
    ) -> AsyncIterator[ChatMessage[TResult]]:
        """Run agent sequentially on multiple prompt groups.

        Args:
            prompt_groups: Groups of prompts to process sequentially
            result_type: Optional type for structured responses
            model: Optional model override
            store_history: Whether to store in conversation history
            wait_for_connections: Whether to wait for connected agents

        Yields:
            Response messages in sequence

        Example:
            questions = [
                ["What is your name?"],
                ["How old are you?", image1],
                ["Describe this image", image2],
            ]
            async for response in agent.run_iter(*questions):
                print(response.content)
        """
        for prompts in prompt_groups:
            response = await self.run(
                *prompts,
                result_type=result_type,
                model=model,
                store_history=store_history,
                wait_for_connections=wait_for_connections,
            )
            yield response  # pyright: ignore

    def run_sync(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        result_type: type[TResult] | None = None,
        deps: TDeps | None = None,
        model: ModelType = None,
        store_history: bool = True,
    ) -> ChatMessage[TResult]:
        """Run agent synchronously (convenience wrapper).

        Args:
            prompt: User query or instruction
            result_type: Optional type for structured responses
            deps: Optional dependencies for the agent
            model: Optional model override
            store_history: Whether the message exchange should be added to the
                           context window
        Returns:
            Result containing response and run information
        """
        coro = self.run(
            *prompt,
            model=model,
            store_history=store_history,
            result_type=result_type,
        )
        return self.run_task_sync(coro)  # type: ignore

    async def run_job(
        self,
        job: Job[TDeps, str | None],
        *,
        store_history: bool = True,
        include_agent_tools: bool = True,
    ) -> ChatMessage[str]:
        """Execute a pre-defined task.

        Args:
            job: Job configuration to execute
            store_history: Whether the message exchange should be added to the
                           context window
            include_agent_tools: Whether to include agent tools
        Returns:
            Job execution result

        Raises:
            JobError: If task execution fails
            ValueError: If task configuration is invalid
        """
        from llmling_agent.tasks import JobError

        if job.required_dependency is not None:  # noqa: SIM102
            if not isinstance(self.context.data, job.required_dependency):
                msg = (
                    f"Agent dependencies ({type(self.context.data)}) "
                    f"don't match job requirement ({job.required_dependency})"
                )
                raise JobError(msg)

        # Load task knowledge
        if job.knowledge:
            # Add knowledge sources to context
            resources: list[Resource | str] = list(job.knowledge.paths) + list(
                job.knowledge.resources
            )
            for source in resources:
                await self.conversation.load_context_source(source)
            for prompt in job.knowledge.prompts:
                await self.conversation.load_context_source(prompt)
        try:
            # Register task tools temporarily
            tools = job.get_tools()
            with self.tools.temporary_tools(tools, exclusive=not include_agent_tools):
                # Execute job with job-specific tools
                return await self.run(await job.get_prompt(), store_history=store_history)

        except Exception as e:
            msg = f"Task execution failed: {e}"
            logger.exception(msg)
            raise JobError(msg) from e

    @asynccontextmanager
    @track_action("Calling Agent.iterate_run: {prompts}")
    async def iterate_run(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        message_id: str | None = None,
        message_history: list[ChatMessage[Any]] | None = None,
        tools: list[Tool] | None = None,
        result_type: type[TResult] | None = None,
        usage_limits: UsageLimits | None = None,
        model: ModelType = None,
        system_prompt: str | None = None,
        tool_choice: str | list[str] | None = None,
        conversation_id: str | None = None,
    ) -> AsyncIterator[AgentRunProtocol[TResult]]:
        """Run the agent step-by-step, yielding an object to observe the execution graph.

        Args:
            *prompts: User query/instructions (text, images, paths, BasePrompts).
            message_id: Optional unique ID for this run attempt. Generated if None.
            message_history: Optional list of messages to replace current history.
            tools: Optional sequence of tools to use instead of agent's default tools.
            result_type: Optional type for structured responses.
            usage_limits: Optional usage limits (provider support may vary).
            model: Optional model override for this run.
            system_prompt: Optional system prompt override for this run.
            tool_choice: Filter agent's tools by name (ignored if `tools` is provided).
            conversation_id: Optional ID to associate with the conversation context.

        Yields:
            An object conforming to AgentRunProtocol for iterating over execution nodes.

        Example: (Same as before)
            async with agent.iterate_run("Capital of France?") as agent_run:
                async for node in agent_run:
                    print(f"Processing: {type(node).__name__}")
                print(f"Final result: {agent_run.result.output}")

        Note: (Same as before regarding history management)
        """
        run_message_id = message_id or str(uuid4())
        logger.info("Starting agent iteration run_id=%s", run_message_id)
        converted_prompts = await convert_prompts(prompts)
        if not converted_prompts:
            msg = "No prompts provided for iteration."
            logger.error(msg)
            raise ValueError(msg)

        if tools is None:
            effective_tools = await self.tools.get_tools(
                state="enabled", names=tool_choice
            )
        else:
            effective_tools = tools  # Use the direct override

        self.set_result_type(result_type)
        effective_system_prompt = (
            system_prompt
            if system_prompt is not None
            else await self.sys_prompts.format_system_prompt(self)
        )
        effective_message_history = (
            message_history
            if message_history is not None
            else self.conversation.get_history()
        )
        try:
            async with self._provider.iterate_run(
                *converted_prompts,
                # Pass consistent arguments to the provider
                message_id=run_message_id,
                message_history=effective_message_history,
                tools=effective_tools,
                result_type=result_type,
                usage_limits=usage_limits,
                model=model,
                system_prompt=effective_system_prompt,
            ) as agent_run_protocol_object:
                yield agent_run_protocol_object

            logger.info("Agent iteration run_id=%s completed.", run_message_id)
            # History update logic could go here if desired

        except Exception as e:
            logger.exception("Agent iteration run_id=%s failed.", run_message_id)
            self.run_failed.emit(f"Agent iteration failed: {e}", e)
            raise

    async def run_in_background(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        max_count: int | None = None,
        interval: float = 1.0,
        block: bool = False,
        **kwargs: Any,
    ) -> ChatMessage[TResult] | None:
        """Run agent continuously in background with prompt or dynamic prompt function.

        Args:
            prompt: Static prompt or function that generates prompts
            max_count: Maximum number of runs (None = infinite)
            interval: Seconds between runs
            block: Whether to block until completion
            **kwargs: Arguments passed to run()
        """
        self._infinite = max_count is None

        async def _continuous():
            count = 0
            msg = "%s: Starting continuous run (max_count=%s, interval=%s) for %r"
            logger.debug(msg, self.name, max_count, interval, self.name)
            latest = None
            while max_count is None or count < max_count:
                try:
                    current_prompts = [
                        call_with_context(p, self.context, **kwargs) if callable(p) else p
                        for p in prompt
                    ]
                    msg = "%s: Generated prompt #%d: %s"
                    logger.debug(msg, self.name, count, current_prompts)

                    latest = await self.run(current_prompts, **kwargs)
                    msg = "%s: Run continous result #%d"
                    logger.debug(msg, self.name, count)

                    count += 1
                    await asyncio.sleep(interval)
                except asyncio.CancelledError:
                    logger.debug("%s: Continuous run cancelled", self.name)
                    break
                except Exception:
                    logger.exception("%s: Background run failed", self.name)
                    await asyncio.sleep(interval)
            msg = "%s: Continuous run completed after %d iterations"
            logger.debug(msg, self.name, count)
            return latest

        # Cancel any existing background task
        await self.stop()
        task = asyncio.create_task(_continuous(), name=f"background_{self.name}")
        if block:
            try:
                return await task  # type: ignore
            finally:
                if not task.done():
                    task.cancel()
        else:
            logger.debug("%s: Started background task %s", self.name, task.get_name())
            self._background_task = task
            return None

    async def stop(self):
        """Stop continuous execution if running."""
        if self._background_task and not self._background_task.done():
            self._background_task.cancel()
            await self._background_task
            self._background_task = None

    async def wait(self) -> ChatMessage[TResult]:
        """Wait for background execution to complete."""
        if not self._background_task:
            msg = "No background task running"
            raise RuntimeError(msg)
        if self._infinite:
            msg = "Cannot wait on infinite execution"
            raise RuntimeError(msg)
        try:
            return await self._background_task
        finally:
            self._background_task = None

    def clear_history(self):
        """Clear both internal and pydantic-ai history."""
        self._logger.clear_state()
        self.conversation.clear()
        logger.debug("Cleared history and reset tool state")

    async def share(
        self,
        target: AnyAgent[TDeps, Any],
        *,
        tools: list[str] | None = None,
        resources: list[str] | None = None,
        history: bool | int | None = None,  # bool or number of messages
        token_limit: int | None = None,
    ):
        """Share capabilities and knowledge with another agent.

        Args:
            target: Agent to share with
            tools: List of tool names to share
            resources: List of resource names to share
            history: Share conversation history:
                    - True: Share full history
                    - int: Number of most recent messages to share
                    - None: Don't share history
            token_limit: Optional max tokens for history

        Raises:
            ValueError: If requested items don't exist
            RuntimeError: If runtime not available for resources
        """
        # Share tools if requested
        for name in tools or []:
            if tool := self.tools.get(name):
                meta = {"shared_from": self.name}
                target.tools.register_tool(tool.callable, metadata=meta)
            else:
                msg = f"Tool not found: {name}"
                raise ValueError(msg)

        # Share resources if requested
        if resources:
            if not self.runtime:
                msg = "No runtime available for sharing resources"
                raise RuntimeError(msg)
            for name in resources:
                if resource := self.runtime.get_resource(name):
                    await target.conversation.load_context_source(resource)  # type: ignore
                else:
                    msg = f"Resource not found: {name}"
                    raise ValueError(msg)

        # Share history if requested
        if history:
            history_text = await self.conversation.format_history(
                max_tokens=token_limit,
                num_messages=history if isinstance(history, int) else None,
            )
            target.conversation.add_context_message(
                history_text, source=self.name, metadata={"type": "shared_history"}
            )

    def register_worker(
        self,
        worker: MessageNode[Any, Any],
        *,
        name: str | None = None,
        reset_history_on_run: bool = True,
        pass_message_history: bool = False,
        share_context: bool = False,
    ) -> Tool:
        """Register another agent as a worker tool."""
        return self.tools.register_worker(
            worker,
            name=name,
            reset_history_on_run=reset_history_on_run,
            pass_message_history=pass_message_history,
            share_context=share_context,
            parent=self if (pass_message_history or share_context) else None,
        )

    def set_model(self, model: ModelType):
        """Set the model for this agent.

        Args:
            model: New model to use (name or instance)

        Emits:
            model_changed signal with the new model
        """
        self._provider.set_model(model)

    async def reset(self):
        """Reset agent state (conversation history and tool states)."""
        old_tools = await self.tools.list_tools()
        self.conversation.clear()
        self.tools.reset_states()
        new_tools = await self.tools.list_tools()

        event = self.AgentReset(
            agent_name=self.name,
            previous_tools=old_tools,
            new_tools=new_tools,
        )
        self.agent_reset.emit(event)

    @property
    def runtime(self) -> RuntimeConfig:
        """Get runtime configuration from context."""
        assert self.context.runtime
        return self.context.runtime

    @runtime.setter
    def runtime(self, value: RuntimeConfig):
        """Set runtime configuration and update context."""
        self.context.runtime = value

    @property
    def stats(self) -> MessageStats:
        return MessageStats(messages=self._logger.message_history)

    @asynccontextmanager
    async def temporary_state(
        self,
        *,
        system_prompts: list[AnyPromptType] | None = None,
        replace_prompts: bool = False,
        tools: list[ToolType] | None = None,
        replace_tools: bool = False,
        history: list[AnyPromptType] | SessionQuery | None = None,
        replace_history: bool = False,
        pause_routing: bool = False,
        model: ModelType | None = None,
        provider: AgentProvider | None = None,
    ) -> AsyncIterator[Self]:
        """Temporarily modify agent state.

        Args:
            system_prompts: Temporary system prompts to use
            replace_prompts: Whether to replace existing prompts
            tools: Temporary tools to make available
            replace_tools: Whether to replace existing tools
            history: Conversation history (prompts or query)
            replace_history: Whether to replace existing history
            pause_routing: Whether to pause message routing
            model: Temporary model override
            provider: Temporary provider override
        """
        old_model = self._provider.model if hasattr(self._provider, "model") else None  # pyright: ignore
        old_provider = self._provider

        async with AsyncExitStack() as stack:
            # System prompts (async)
            if system_prompts is not None:
                await stack.enter_async_context(
                    self.sys_prompts.temporary_prompt(
                        system_prompts, exclusive=replace_prompts
                    )
                )

            # Tools (sync)
            if tools is not None:
                stack.enter_context(
                    self.tools.temporary_tools(tools, exclusive=replace_tools)
                )

            # History (async)
            if history is not None:
                await stack.enter_async_context(
                    self.conversation.temporary_state(
                        history, replace_history=replace_history
                    )
                )

            # Routing (async)
            if pause_routing:
                await stack.enter_async_context(self.connections.paused_routing())

            # Model/Provider
            if provider is not None:
                self._provider = provider
            elif model is not None:
                self._provider.set_model(model)

            try:
                yield self
            finally:
                # Restore model/provider
                if provider is not None:
                    self._provider = old_provider
                elif model is not None and old_model:
                    self._provider.set_model(old_model)

context property writable

context: AgentContext[TDeps]

Get agent context.

model_name property

model_name: str | None

Get the model name in a consistent format.

name property writable

name: str

Get agent name.

provider property writable

provider: AgentProvider

Get the underlying provider.

runtime property writable

runtime: RuntimeConfig

Get runtime configuration from context.

AgentReset dataclass

Emitted when agent is reset.

Source code in src/llmling_agent/agent/agent.py

@dataclass(frozen=True)
class AgentReset:
    """Emitted when agent is reset."""

    agent_name: AgentName
    previous_tools: dict[str, bool]
    new_tools: dict[str, bool]
    timestamp: datetime = field(default_factory=get_now)

__aenter__ async

__aenter__() -> Self

Enter async context and set up MCP servers.

Source code in src/llmling_agent/agent/agent.py

async def __aenter__(self) -> Self:
    """Enter async context and set up MCP servers."""
    try:
        # Collect all coroutines that need to be run
        coros: list[Coroutine[Any, Any, Any]] = []

        # Runtime initialization if needed
        runtime_ref = self.context.runtime
        if runtime_ref and not runtime_ref._initialized:
            self._owns_runtime = True
            coros.append(runtime_ref.__aenter__())

        # Events initialization
        coros.append(super().__aenter__())

        # Get conversation init tasks directly
        coros.extend(self.conversation.get_initialization_tasks())

        # Execute coroutines either in parallel or sequentially
        if self.parallel_init and coros:
            await asyncio.gather(*coros)
        else:
            for coro in coros:
                await coro
        if runtime_ref:
            self.tools.add_provider(RuntimeResourceProvider(runtime_ref))
        for provider in await self.context.config.get_toolsets():
            self.tools.add_provider(provider)
    except Exception as e:
        # Clean up in reverse order
        if self._owns_runtime and runtime_ref and self.context.runtime == runtime_ref:
            await runtime_ref.__aexit__(type(e), e, e.__traceback__)
        msg = "Failed to initialize agent"
        raise RuntimeError(msg) from e
    else:
        return self

__aexit__ async

__aexit__(
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
)

Exit async context.

Source code in src/llmling_agent/agent/agent.py

async def __aexit__(
    self,
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
):
    """Exit async context."""
    await super().__aexit__(exc_type, exc_val, exc_tb)
    try:
        await self.mcp.__aexit__(exc_type, exc_val, exc_tb)
    finally:
        if self._owns_runtime and self.context.runtime:
            self.tools.remove_provider("runtime")
            await self.context.runtime.__aexit__(exc_type, exc_val, exc_tb)

__and__

__and__(other: Agent[TDeps] | StructuredAgent[TDeps, Any]) -> Team[TDeps]

__and__(other: Team[TDeps]) -> Team[TDeps]

__and__(other: ProcessorCallback[Any]) -> Team[TDeps]

__and__(other: MessageNode[Any, Any] | ProcessorCallback[Any]) -> Team[Any]

Create agent group using | operator.

Example

group = analyzer & planner & executor # Create group of 3 group = analyzer & existing_group # Add to existing group

Source code in src/llmling_agent/agent/agent.py

def __and__(self, other: MessageNode[Any, Any] | ProcessorCallback[Any]) -> Team[Any]:
    """Create agent group using | operator.

    Example:
        group = analyzer & planner & executor  # Create group of 3
        group = analyzer & existing_group  # Add to existing group
    """
    from llmling_agent.agent import StructuredAgent
    from llmling_agent.delegation.team import Team

    match other:
        case Team():
            return Team([self, *other.agents])
        case Callable():
            if callable(other):
                if has_return_type(other, str):
                    agent_2 = Agent.from_callback(other)
                else:
                    agent_2 = StructuredAgent.from_callback(other)
            agent_2.context.pool = self.context.pool
            return Team([self, agent_2])
        case MessageNode():
            return Team([self, other])
        case _:
            msg = f"Invalid agent type: {type(other)}"
            raise ValueError(msg)

__init__

__init__(
    name: str = "llmling-agent",
    provider: AgentType = "pydantic_ai",
    *,
    model: ModelType = None,
    runtime: RuntimeConfig | Config | StrPath | None = None,
    context: AgentContext[TDeps] | None = None,
    session: SessionIdType | SessionQuery | MemoryConfig | bool | int = None,
    system_prompt: AnyPromptType | Sequence[AnyPromptType] = (),
    description: str | None = None,
    tools: Sequence[ToolType] | None = None,
    capabilities: Capabilities | None = None,
    mcp_servers: Sequence[str | MCPServerConfig] | None = None,
    resources: Sequence[Resource | PromptType | str] = (),
    retries: int = 1,
    result_retries: int | None = None,
    end_strategy: EndStrategy = "early",
    defer_model_check: bool = False,
    input_provider: InputProvider | None = None,
    parallel_init: bool = True,
    debug: bool = False,
)

Initialize agent with runtime configuration.

Parameters:

Name	Type	Description	Default
runtime	RuntimeConfig | Config | StrPath | None	Runtime configuration providing access to resources/tools	None
context	AgentContext[TDeps] | None	Agent context with capabilities and configuration	None
provider	AgentType	Agent type to use (ai: PydanticAIProvider, human: HumanProvider)	'pydantic_ai'
session	SessionIdType | SessionQuery | MemoryConfig | bool | int	Memory configuration. - None: Default memory config - False: Disable message history (max_messages=0) - int: Max tokens for memory - str/UUID: Session identifier - SessionQuery: Query to recover conversation - MemoryConfig: Complete memory configuration	None
model	ModelType	The default model to use (defaults to GPT-4)	None
system_prompt	AnyPromptType | Sequence[AnyPromptType]	Static system prompts to use for this agent	()
name	str	Name of the agent for logging	'llmling-agent'
description	str | None	Description of the Agent ("what it can do")	None
tools	Sequence[ToolType] | None	List of tools to register with the agent	None
capabilities	Capabilities | None	Capabilities for the agent	None
mcp_servers	Sequence[str | MCPServerConfig] | None	MCP servers to connect to	None
resources	Sequence[Resource | PromptType | str]	Additional resources to load	()
retries	int	Default number of retries for failed operations	1
result_retries	int | None	Max retries for result validation (defaults to retries)	None
end_strategy	EndStrategy	Strategy for handling tool calls that are requested alongside a final result	'early'
defer_model_check	bool	Whether to defer model evaluation until first run	False
input_provider	InputProvider | None	Provider for human input (tool confirmation / HumanProviders)	None
parallel_init	bool	Whether to initialize resources in parallel	True
debug	bool	Whether to enable debug mode	False

Source code in src/llmling_agent/agent/agent.py

def __init__(  # noqa: PLR0915
    # we dont use AgentKwargs here so that we can work with explicit ones in the ctor
    self,
    name: str = "llmling-agent",
    provider: AgentType = "pydantic_ai",
    *,
    model: ModelType = None,
    runtime: RuntimeConfig | Config | StrPath | None = None,
    context: AgentContext[TDeps] | None = None,
    session: SessionIdType | SessionQuery | MemoryConfig | bool | int = None,
    system_prompt: AnyPromptType | Sequence[AnyPromptType] = (),
    description: str | None = None,
    tools: Sequence[ToolType] | None = None,
    capabilities: Capabilities | None = None,
    mcp_servers: Sequence[str | MCPServerConfig] | None = None,
    resources: Sequence[Resource | PromptType | str] = (),
    retries: int = 1,
    result_retries: int | None = None,
    end_strategy: EndStrategy = "early",
    defer_model_check: bool = False,
    input_provider: InputProvider | None = None,
    parallel_init: bool = True,
    debug: bool = False,
):
    """Initialize agent with runtime configuration.

    Args:
        runtime: Runtime configuration providing access to resources/tools
        context: Agent context with capabilities and configuration
        provider: Agent type to use (ai: PydanticAIProvider, human: HumanProvider)
        session: Memory configuration.
            - None: Default memory config
            - False: Disable message history (max_messages=0)
            - int: Max tokens for memory
            - str/UUID: Session identifier
            - SessionQuery: Query to recover conversation
            - MemoryConfig: Complete memory configuration
        model: The default model to use (defaults to GPT-4)
        system_prompt: Static system prompts to use for this agent
        name: Name of the agent for logging
        description: Description of the Agent ("what it can do")
        tools: List of tools to register with the agent
        capabilities: Capabilities for the agent
        mcp_servers: MCP servers to connect to
        resources: Additional resources to load
        retries: Default number of retries for failed operations
        result_retries: Max retries for result validation (defaults to retries)
        end_strategy: Strategy for handling tool calls that are requested alongside
                      a final result
        defer_model_check: Whether to defer model evaluation until first run
        input_provider: Provider for human input (tool confirmation / HumanProviders)
        parallel_init: Whether to initialize resources in parallel
        debug: Whether to enable debug mode
    """
    from llmling_agent.agent import AgentContext
    from llmling_agent.agent.conversation import ConversationManager
    from llmling_agent.agent.interactions import Interactions
    from llmling_agent.agent.sys_prompts import SystemPrompts
    from llmling_agent.resource_providers.capability_provider import (
        CapabilitiesResourceProvider,
    )
    from llmling_agent_providers.base import AgentProvider

    self._infinite = False
    # save some stuff for asnyc init
    self._owns_runtime = False
    # prepare context
    ctx = context or AgentContext[TDeps].create_default(
        name,
        input_provider=input_provider,
        capabilities=capabilities,
    )
    self._context = ctx
    memory_cfg = (
        session
        if isinstance(session, MemoryConfig)
        else MemoryConfig.from_value(session)
    )
    super().__init__(
        name=name,
        context=ctx,
        description=description,
        enable_logging=memory_cfg.enable,
        mcp_servers=mcp_servers,
    )
    # Initialize runtime
    match runtime:
        case None:
            ctx.runtime = RuntimeConfig.from_config(Config())
        case Config() | str() | PathLike():
            ctx.runtime = RuntimeConfig.from_config(runtime)
        case RuntimeConfig():
            ctx.runtime = runtime

    runtime_provider = RuntimePromptProvider(ctx.runtime)
    ctx.definition.prompt_manager.providers["runtime"] = runtime_provider
    # Initialize tool manager
    all_tools = list(tools or [])
    self.tools = ToolManager(all_tools)
    self.tools.add_provider(self.mcp)
    if builtin_tools := ctx.config.get_tool_provider():
        self.tools.add_provider(builtin_tools)

    # Initialize conversation manager
    resources = list(resources)
    if ctx.config.knowledge:
        resources.extend(ctx.config.knowledge.get_resources())
    self.conversation = ConversationManager(self, memory_cfg, resources=resources)
    # Initialize provider
    match provider:
        case "pydantic_ai":
            validate_import("pydantic_ai", "pydantic_ai")
            from llmling_agent_providers.pydanticai import PydanticAIProvider

            if model and not isinstance(model, str):
                from pydantic_ai import models

                assert isinstance(model, models.Model)
            self._provider: AgentProvider = PydanticAIProvider(
                model=model,
                retries=retries,
                end_strategy=end_strategy,
                result_retries=result_retries,
                defer_model_check=defer_model_check,
                debug=debug,
                context=ctx,
            )
        case "human":
            from llmling_agent_providers.human import HumanProvider

            self._provider = HumanProvider(name=name, debug=debug, context=ctx)
        case Callable():
            from llmling_agent_providers.callback import CallbackProvider

            self._provider = CallbackProvider(
                provider, name=name, debug=debug, context=ctx
            )
        case "litellm":
            validate_import("litellm", "litellm")
            from llmling_agent_providers.litellm_provider import LiteLLMProvider

            self._provider = LiteLLMProvider(
                name=name,
                debug=debug,
                retries=retries,
                context=ctx,
                model=model,
            )
        case AgentProvider():
            self._provider = provider
            self._provider.context = ctx
        case _:
            msg = f"Invalid agent type: {type}"
            raise ValueError(msg)
    self.tools.add_provider(CapabilitiesResourceProvider(ctx.capabilities))

    if ctx and ctx.definition:
        from llmling_agent.observability import registry

        registry.register_providers(ctx.definition.observability)

    # init variables
    self._debug = debug
    self._result_type: type | None = None
    self.parallel_init = parallel_init
    self.name = name
    self._background_task: asyncio.Task[Any] | None = None

    # Forward provider signals
    self._provider.chunk_streamed.connect(self.chunk_streamed)
    self._provider.model_changed.connect(self.model_changed)
    self._provider.tool_used.connect(self.tool_used)
    self._provider.model_changed.connect(self.model_changed)

    self.talk = Interactions(self)

    # Set up system prompts
    config_prompts = ctx.config.system_prompts if ctx else []
    all_prompts: list[AnyPromptType] = list(config_prompts)
    if isinstance(system_prompt, list):
        all_prompts.extend(system_prompt)
    else:
        all_prompts.append(system_prompt)
    self.sys_prompts = SystemPrompts(all_prompts, context=ctx)

clear_history

clear_history()

Clear both internal and pydantic-ai history.

Source code in src/llmling_agent/agent/agent.py

def clear_history(self):
    """Clear both internal and pydantic-ai history."""
    self._logger.clear_state()
    self.conversation.clear()
    logger.debug("Cleared history and reset tool state")

from_callback classmethod

from_callback(
    callback: ProcessorCallback[str],
    *,
    name: str | None = None,
    debug: bool = False,
    **kwargs: Any,
) -> Agent[None]

Create an agent from a processing callback.

Parameters:

Name	Type	Description	Default
callback	ProcessorCallback[str]	Function to process messages. Can be: - sync or async - with or without context - must return str for pipeline compatibility	required
name	str | None	Optional name for the agent	None
debug	bool	Whether to enable debug mode	False
kwargs	Any	Additional arguments for agent	{}

Source code in src/llmling_agent/agent/agent.py

@classmethod
def from_callback(
    cls,
    callback: ProcessorCallback[str],
    *,
    name: str | None = None,
    debug: bool = False,
    **kwargs: Any,
) -> Agent[None]:
    """Create an agent from a processing callback.

    Args:
        callback: Function to process messages. Can be:
            - sync or async
            - with or without context
            - must return str for pipeline compatibility
        name: Optional name for the agent
        debug: Whether to enable debug mode
        kwargs: Additional arguments for agent
    """
    from llmling_agent_providers.callback import CallbackProvider

    name = name or callback.__name__ or "processor"
    provider = CallbackProvider(callback, name=name)
    return Agent[None](provider=provider, name=name, debug=debug, **kwargs)

is_busy

is_busy() -> bool

Check if agent is currently processing tasks.

Source code in src/llmling_agent/agent/agent.py

def is_busy(self) -> bool:
    """Check if agent is currently processing tasks."""
    return bool(self._pending_tasks or self._background_task)

iterate_run async

iterate_run(
    *prompts: AnyPromptType | Image | PathLike[str],
    message_id: str | None = None,
    message_history: list[ChatMessage[Any]] | None = None,
    tools: list[Tool] | None = None,
    result_type: type[TResult] | None = None,
    usage_limits: UsageLimits | None = None,
    model: ModelType = None,
    system_prompt: str | None = None,
    tool_choice: str | list[str] | None = None,
    conversation_id: str | None = None,
) -> AsyncIterator[AgentRunProtocol[TResult]]

Run the agent step-by-step, yielding an object to observe the execution graph.

Parameters:

Name	Type	Description	Default
*prompts	AnyPromptType | Image | PathLike[str]	User query/instructions (text, images, paths, BasePrompts).	()
message_id	str | None	Optional unique ID for this run attempt. Generated if None.	None
message_history	list[ChatMessage[Any]] | None	Optional list of messages to replace current history.	None
tools	list[Tool] | None	Optional sequence of tools to use instead of agent's default tools.	None
result_type	type[TResult] | None	Optional type for structured responses.	None
usage_limits	UsageLimits | None	Optional usage limits (provider support may vary).	None
model	ModelType	Optional model override for this run.	None
system_prompt	str | None	Optional system prompt override for this run.	None
tool_choice	str | list[str] | None	Filter agent's tools by name (ignored if tools is provided).	None
conversation_id	str | None	Optional ID to associate with the conversation context.	None

Yields:

Type	Description
AsyncIterator[AgentRunProtocol[TResult]]	An object conforming to AgentRunProtocol for iterating over execution nodes.

(Same as before)

async with agent.iterate_run("Capital of France?") as agent_run: async for node in agent_run: print(f"Processing: {type(node).name}") print(f"Final result: {agent_run.result.output}")

Note: (Same as before regarding history management)

Source code in src/llmling_agent/agent/agent.py

@asynccontextmanager
@track_action("Calling Agent.iterate_run: {prompts}")
async def iterate_run(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    message_id: str | None = None,
    message_history: list[ChatMessage[Any]] | None = None,
    tools: list[Tool] | None = None,
    result_type: type[TResult] | None = None,
    usage_limits: UsageLimits | None = None,
    model: ModelType = None,
    system_prompt: str | None = None,
    tool_choice: str | list[str] | None = None,
    conversation_id: str | None = None,
) -> AsyncIterator[AgentRunProtocol[TResult]]:
    """Run the agent step-by-step, yielding an object to observe the execution graph.

    Args:
        *prompts: User query/instructions (text, images, paths, BasePrompts).
        message_id: Optional unique ID for this run attempt. Generated if None.
        message_history: Optional list of messages to replace current history.
        tools: Optional sequence of tools to use instead of agent's default tools.
        result_type: Optional type for structured responses.
        usage_limits: Optional usage limits (provider support may vary).
        model: Optional model override for this run.
        system_prompt: Optional system prompt override for this run.
        tool_choice: Filter agent's tools by name (ignored if `tools` is provided).
        conversation_id: Optional ID to associate with the conversation context.

    Yields:
        An object conforming to AgentRunProtocol for iterating over execution nodes.

    Example: (Same as before)
        async with agent.iterate_run("Capital of France?") as agent_run:
            async for node in agent_run:
                print(f"Processing: {type(node).__name__}")
            print(f"Final result: {agent_run.result.output}")

    Note: (Same as before regarding history management)
    """
    run_message_id = message_id or str(uuid4())
    logger.info("Starting agent iteration run_id=%s", run_message_id)
    converted_prompts = await convert_prompts(prompts)
    if not converted_prompts:
        msg = "No prompts provided for iteration."
        logger.error(msg)
        raise ValueError(msg)

    if tools is None:
        effective_tools = await self.tools.get_tools(
            state="enabled", names=tool_choice
        )
    else:
        effective_tools = tools  # Use the direct override

    self.set_result_type(result_type)
    effective_system_prompt = (
        system_prompt
        if system_prompt is not None
        else await self.sys_prompts.format_system_prompt(self)
    )
    effective_message_history = (
        message_history
        if message_history is not None
        else self.conversation.get_history()
    )
    try:
        async with self._provider.iterate_run(
            *converted_prompts,
            # Pass consistent arguments to the provider
            message_id=run_message_id,
            message_history=effective_message_history,
            tools=effective_tools,
            result_type=result_type,
            usage_limits=usage_limits,
            model=model,
            system_prompt=effective_system_prompt,
        ) as agent_run_protocol_object:
            yield agent_run_protocol_object

        logger.info("Agent iteration run_id=%s completed.", run_message_id)
        # History update logic could go here if desired

    except Exception as e:
        logger.exception("Agent iteration run_id=%s failed.", run_message_id)
        self.run_failed.emit(f"Agent iteration failed: {e}", e)
        raise

register_worker

register_worker(
    worker: MessageNode[Any, Any],
    *,
    name: str | None = None,
    reset_history_on_run: bool = True,
    pass_message_history: bool = False,
    share_context: bool = False,
) -> Tool

Register another agent as a worker tool.

Source code in src/llmling_agent/agent/agent.py

def register_worker(
    self,
    worker: MessageNode[Any, Any],
    *,
    name: str | None = None,
    reset_history_on_run: bool = True,
    pass_message_history: bool = False,
    share_context: bool = False,
) -> Tool:
    """Register another agent as a worker tool."""
    return self.tools.register_worker(
        worker,
        name=name,
        reset_history_on_run=reset_history_on_run,
        pass_message_history=pass_message_history,
        share_context=share_context,
        parent=self if (pass_message_history or share_context) else None,
    )

reset async

reset()

Reset agent state (conversation history and tool states).

Source code in src/llmling_agent/agent/agent.py

async def reset(self):
    """Reset agent state (conversation history and tool states)."""
    old_tools = await self.tools.list_tools()
    self.conversation.clear()
    self.tools.reset_states()
    new_tools = await self.tools.list_tools()

    event = self.AgentReset(
        agent_name=self.name,
        previous_tools=old_tools,
        new_tools=new_tools,
    )
    self.agent_reset.emit(event)

run_in_background async

run_in_background(
    *prompt: AnyPromptType | Image | PathLike[str],
    max_count: int | None = None,
    interval: float = 1.0,
    block: bool = False,
    **kwargs: Any,
) -> ChatMessage[TResult] | None

Run agent continuously in background with prompt or dynamic prompt function.

Parameters:

Name	Type	Description	Default
prompt	AnyPromptType | Image | PathLike[str]	Static prompt or function that generates prompts	()
max_count	int | None	Maximum number of runs (None = infinite)	None
interval	float	Seconds between runs	1.0
block	bool	Whether to block until completion	False
**kwargs	Any	Arguments passed to run()	{}

Source code in src/llmling_agent/agent/agent.py

async def run_in_background(
    self,
    *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    max_count: int | None = None,
    interval: float = 1.0,
    block: bool = False,
    **kwargs: Any,
) -> ChatMessage[TResult] | None:
    """Run agent continuously in background with prompt or dynamic prompt function.

    Args:
        prompt: Static prompt or function that generates prompts
        max_count: Maximum number of runs (None = infinite)
        interval: Seconds between runs
        block: Whether to block until completion
        **kwargs: Arguments passed to run()
    """
    self._infinite = max_count is None

    async def _continuous():
        count = 0
        msg = "%s: Starting continuous run (max_count=%s, interval=%s) for %r"
        logger.debug(msg, self.name, max_count, interval, self.name)
        latest = None
        while max_count is None or count < max_count:
            try:
                current_prompts = [
                    call_with_context(p, self.context, **kwargs) if callable(p) else p
                    for p in prompt
                ]
                msg = "%s: Generated prompt #%d: %s"
                logger.debug(msg, self.name, count, current_prompts)

                latest = await self.run(current_prompts, **kwargs)
                msg = "%s: Run continous result #%d"
                logger.debug(msg, self.name, count)

                count += 1
                await asyncio.sleep(interval)
            except asyncio.CancelledError:
                logger.debug("%s: Continuous run cancelled", self.name)
                break
            except Exception:
                logger.exception("%s: Background run failed", self.name)
                await asyncio.sleep(interval)
        msg = "%s: Continuous run completed after %d iterations"
        logger.debug(msg, self.name, count)
        return latest

    # Cancel any existing background task
    await self.stop()
    task = asyncio.create_task(_continuous(), name=f"background_{self.name}")
    if block:
        try:
            return await task  # type: ignore
        finally:
            if not task.done():
                task.cancel()
    else:
        logger.debug("%s: Started background task %s", self.name, task.get_name())
        self._background_task = task
        return None

run_iter async

run_iter(
    *prompt_groups: Sequence[AnyPromptType | Image | PathLike[str]],
    result_type: type[TResult] | None = None,
    model: ModelType = None,
    store_history: bool = True,
    wait_for_connections: bool | None = None,
) -> AsyncIterator[ChatMessage[TResult]]

Run agent sequentially on multiple prompt groups.

Parameters:

Name	Type	Description	Default
prompt_groups	Sequence[AnyPromptType | Image | PathLike[str]]	Groups of prompts to process sequentially	()
result_type	type[TResult] | None	Optional type for structured responses	None
model	ModelType	Optional model override	None
store_history	bool	Whether to store in conversation history	True
wait_for_connections	bool | None	Whether to wait for connected agents	None

Yields:

Type	Description
AsyncIterator[ChatMessage[TResult]]	Response messages in sequence

Example

questions = [ ["What is your name?"], ["How old are you?", image1], ["Describe this image", image2], ] async for response in agent.run_iter(*questions): print(response.content)

Source code in src/llmling_agent/agent/agent.py

async def run_iter(
    self,
    *prompt_groups: Sequence[AnyPromptType | PIL.Image.Image | os.PathLike[str]],
    result_type: type[TResult] | None = None,
    model: ModelType = None,
    store_history: bool = True,
    wait_for_connections: bool | None = None,
) -> AsyncIterator[ChatMessage[TResult]]:
    """Run agent sequentially on multiple prompt groups.

    Args:
        prompt_groups: Groups of prompts to process sequentially
        result_type: Optional type for structured responses
        model: Optional model override
        store_history: Whether to store in conversation history
        wait_for_connections: Whether to wait for connected agents

    Yields:
        Response messages in sequence

    Example:
        questions = [
            ["What is your name?"],
            ["How old are you?", image1],
            ["Describe this image", image2],
        ]
        async for response in agent.run_iter(*questions):
            print(response.content)
    """
    for prompts in prompt_groups:
        response = await self.run(
            *prompts,
            result_type=result_type,
            model=model,
            store_history=store_history,
            wait_for_connections=wait_for_connections,
        )
        yield response  # pyright: ignore

run_job async

run_job(
    job: Job[TDeps, str | None],
    *,
    store_history: bool = True,
    include_agent_tools: bool = True,
) -> ChatMessage[str]

Execute a pre-defined task.

Parameters:

Name	Type	Description	Default
job	Job[TDeps, str | None]	Job configuration to execute	required
store_history	bool	Whether the message exchange should be added to the context window	True
include_agent_tools	bool	Whether to include agent tools	True

Returns: Job execution result

Raises:

Type	Description
JobError	If task execution fails
ValueError	If task configuration is invalid

Source code in src/llmling_agent/agent/agent.py

async def run_job(
    self,
    job: Job[TDeps, str | None],
    *,
    store_history: bool = True,
    include_agent_tools: bool = True,
) -> ChatMessage[str]:
    """Execute a pre-defined task.

    Args:
        job: Job configuration to execute
        store_history: Whether the message exchange should be added to the
                       context window
        include_agent_tools: Whether to include agent tools
    Returns:
        Job execution result

    Raises:
        JobError: If task execution fails
        ValueError: If task configuration is invalid
    """
    from llmling_agent.tasks import JobError

    if job.required_dependency is not None:  # noqa: SIM102
        if not isinstance(self.context.data, job.required_dependency):
            msg = (
                f"Agent dependencies ({type(self.context.data)}) "
                f"don't match job requirement ({job.required_dependency})"
            )
            raise JobError(msg)

    # Load task knowledge
    if job.knowledge:
        # Add knowledge sources to context
        resources: list[Resource | str] = list(job.knowledge.paths) + list(
            job.knowledge.resources
        )
        for source in resources:
            await self.conversation.load_context_source(source)
        for prompt in job.knowledge.prompts:
            await self.conversation.load_context_source(prompt)
    try:
        # Register task tools temporarily
        tools = job.get_tools()
        with self.tools.temporary_tools(tools, exclusive=not include_agent_tools):
            # Execute job with job-specific tools
            return await self.run(await job.get_prompt(), store_history=store_history)

    except Exception as e:
        msg = f"Task execution failed: {e}"
        logger.exception(msg)
        raise JobError(msg) from e

run_stream async

run_stream(
    *prompt: AnyPromptType | Image | PathLike[str],
    result_type: type[TResult] | None = None,
    model: ModelType = None,
    tool_choice: str | list[str] | None = None,
    store_history: bool = True,
    usage_limits: UsageLimits | None = None,
    message_id: str | None = None,
    conversation_id: str | None = None,
    messages: list[ChatMessage[Any]] | None = None,
    wait_for_connections: bool | None = None,
) -> AsyncIterator[StreamingResponseProtocol[TResult]]

Run agent with prompt and get a streaming response.

Parameters:

Name	Type	Description	Default
prompt	AnyPromptType | Image | PathLike[str]	User query or instruction	()
result_type	type[TResult] | None	Optional type for structured responses	None
model	ModelType	Optional model override	None
tool_choice	str | list[str] | None	Filter tool choice by name	None
store_history	bool	Whether the message exchange should be added to the context window	True
usage_limits	UsageLimits | None	Optional usage limits for the model	None
message_id	str | None	Optional message id for the returned message. Automatically generated if not provided.	None
conversation_id	str | None	Optional conversation id for the returned message.	None
messages	list[ChatMessage[Any]] | None	Optional list of messages to replace the conversation history	None
wait_for_connections	bool | None	Whether to wait for connected agents to complete	None

Returns:

Type	Description
AsyncIterator[StreamingResponseProtocol[TResult]]	A streaming result to iterate over.

Raises:

Type	Description
UnexpectedModelBehavior	If the model fails or behaves unexpectedly

Source code in src/llmling_agent/agent/agent.py

@asynccontextmanager
async def run_stream(
    self,
    *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    result_type: type[TResult] | None = None,
    model: ModelType = None,
    tool_choice: str | list[str] | None = None,
    store_history: bool = True,
    usage_limits: UsageLimits | None = None,
    message_id: str | None = None,
    conversation_id: str | None = None,
    messages: list[ChatMessage[Any]] | None = None,
    wait_for_connections: bool | None = None,
) -> AsyncIterator[StreamingResponseProtocol[TResult]]:
    """Run agent with prompt and get a streaming response.

    Args:
        prompt: User query or instruction
        result_type: Optional type for structured responses
        model: Optional model override
        tool_choice: Filter tool choice by name
        store_history: Whether the message exchange should be added to the
                       context window
        usage_limits: Optional usage limits for the model
        message_id: Optional message id for the returned message.
                    Automatically generated if not provided.
        conversation_id: Optional conversation id for the returned message.
        messages: Optional list of messages to replace the conversation history
        wait_for_connections: Whether to wait for connected agents to complete

    Returns:
        A streaming result to iterate over.

    Raises:
        UnexpectedModelBehavior: If the model fails or behaves unexpectedly
    """
    message_id = message_id or str(uuid4())
    user_msg, prompts = await self.pre_run(*prompt)
    self.set_result_type(result_type)
    start_time = time.perf_counter()
    sys_prompt = await self.sys_prompts.format_system_prompt(self)
    tools = await self.tools.get_tools(state="enabled", names=tool_choice)
    message_history = (
        messages if messages is not None else self.conversation.get_history()
    )
    try:
        async with self._provider.stream_response(
            *prompts,
            message_id=message_id,
            message_history=message_history,
            result_type=result_type,
            model=model,
            store_history=store_history,
            tools=tools,
            usage_limits=usage_limits,
            system_prompt=sys_prompt,
        ) as stream:
            yield stream
            usage = stream.usage()
            cost_info = None
            model_name = stream.model_name  # type: ignore
            if model_name:
                cost_info = await TokenCost.from_usage(
                    usage,
                    model_name,
                    str(user_msg.content),
                    str(stream.formatted_content),  # type: ignore
                )
            response_msg = ChatMessage[TResult](
                content=cast(TResult, stream.formatted_content),  # type: ignore
                role="assistant",
                name=self.name,
                model=model_name,
                message_id=message_id,
                conversation_id=user_msg.conversation_id,
                cost_info=cost_info,
                response_time=time.perf_counter() - start_time,
                # provider_extra=stream.provider_extra or {},
            )
            self.message_sent.emit(response_msg)
            if store_history:
                self.conversation.add_chat_messages([user_msg, response_msg])
            await self.connections.route_message(
                response_msg,
                wait=wait_for_connections,
            )

    except Exception as e:
        logger.exception("Agent stream failed")
        self.run_failed.emit("Agent stream failed", e)
        raise

run_sync

run_sync(
    *prompt: AnyPromptType | Image | PathLike[str],
    result_type: type[TResult] | None = None,
    deps: TDeps | None = None,
    model: ModelType = None,
    store_history: bool = True,
) -> ChatMessage[TResult]

Run agent synchronously (convenience wrapper).

Parameters:

Name	Type	Description	Default
prompt	AnyPromptType | Image | PathLike[str]	User query or instruction	()
result_type	type[TResult] | None	Optional type for structured responses	None
deps	TDeps | None	Optional dependencies for the agent	None
model	ModelType	Optional model override	None
store_history	bool	Whether the message exchange should be added to the context window	True

Returns: Result containing response and run information

Source code in src/llmling_agent/agent/agent.py

def run_sync(
    self,
    *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    result_type: type[TResult] | None = None,
    deps: TDeps | None = None,
    model: ModelType = None,
    store_history: bool = True,
) -> ChatMessage[TResult]:
    """Run agent synchronously (convenience wrapper).

    Args:
        prompt: User query or instruction
        result_type: Optional type for structured responses
        deps: Optional dependencies for the agent
        model: Optional model override
        store_history: Whether the message exchange should be added to the
                       context window
    Returns:
        Result containing response and run information
    """
    coro = self.run(
        *prompt,
        model=model,
        store_history=store_history,
        result_type=result_type,
    )
    return self.run_task_sync(coro)  # type: ignore

set_model

set_model(model: ModelType)

Set the model for this agent.

Parameters:

Name	Type	Description	Default
model	ModelType	New model to use (name or instance)	required

Emits

model_changed signal with the new model

Source code in src/llmling_agent/agent/agent.py

def set_model(self, model: ModelType):
    """Set the model for this agent.

    Args:
        model: New model to use (name or instance)

    Emits:
        model_changed signal with the new model
    """
    self._provider.set_model(model)

set_result_type

set_result_type(
    result_type: type[TResult] | str | ResponseDefinition | None,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
)

Set or update the result type for this agent.

Parameters:

Name	Type	Description	Default
result_type	type[TResult] | str | ResponseDefinition | None	New result type, can be: - A Python type for validation - Name of a response definition - Response definition instance - None to reset to unstructured mode	required
tool_name	str | None	Optional override for tool name	None
tool_description	str | None	Optional override for tool description	None

Source code in src/llmling_agent/agent/agent.py

def set_result_type(
    self,
    result_type: type[TResult] | str | ResponseDefinition | None,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
):
    """Set or update the result type for this agent.

    Args:
        result_type: New result type, can be:
            - A Python type for validation
            - Name of a response definition
            - Response definition instance
            - None to reset to unstructured mode
        tool_name: Optional override for tool name
        tool_description: Optional override for tool description
    """
    logger.debug("Setting result type to: %s for %r", result_type, self.name)
    self._result_type = to_type(result_type)

share async

share(
    target: AnyAgent[TDeps, Any],
    *,
    tools: list[str] | None = None,
    resources: list[str] | None = None,
    history: bool | int | None = None,
    token_limit: int | None = None,
)

Share capabilities and knowledge with another agent.

Parameters:

Name	Type	Description	Default
target	AnyAgent[TDeps, Any]	Agent to share with	required
tools	list[str] | None	List of tool names to share	None
resources	list[str] | None	List of resource names to share	None
history	bool | int | None	Share conversation history: - True: Share full history - int: Number of most recent messages to share - None: Don't share history	None
token_limit	int | None	Optional max tokens for history	None

Raises:

Type	Description
ValueError	If requested items don't exist
RuntimeError	If runtime not available for resources

Source code in src/llmling_agent/agent/agent.py

async def share(
    self,
    target: AnyAgent[TDeps, Any],
    *,
    tools: list[str] | None = None,
    resources: list[str] | None = None,
    history: bool | int | None = None,  # bool or number of messages
    token_limit: int | None = None,
):
    """Share capabilities and knowledge with another agent.

    Args:
        target: Agent to share with
        tools: List of tool names to share
        resources: List of resource names to share
        history: Share conversation history:
                - True: Share full history
                - int: Number of most recent messages to share
                - None: Don't share history
        token_limit: Optional max tokens for history

    Raises:
        ValueError: If requested items don't exist
        RuntimeError: If runtime not available for resources
    """
    # Share tools if requested
    for name in tools or []:
        if tool := self.tools.get(name):
            meta = {"shared_from": self.name}
            target.tools.register_tool(tool.callable, metadata=meta)
        else:
            msg = f"Tool not found: {name}"
            raise ValueError(msg)

    # Share resources if requested
    if resources:
        if not self.runtime:
            msg = "No runtime available for sharing resources"
            raise RuntimeError(msg)
        for name in resources:
            if resource := self.runtime.get_resource(name):
                await target.conversation.load_context_source(resource)  # type: ignore
            else:
                msg = f"Resource not found: {name}"
                raise ValueError(msg)

    # Share history if requested
    if history:
        history_text = await self.conversation.format_history(
            max_tokens=token_limit,
            num_messages=history if isinstance(history, int) else None,
        )
        target.conversation.add_context_message(
            history_text, source=self.name, metadata={"type": "shared_history"}
        )

stop async

stop()

Stop continuous execution if running.

Source code in src/llmling_agent/agent/agent.py

async def stop(self):
    """Stop continuous execution if running."""
    if self._background_task and not self._background_task.done():
        self._background_task.cancel()
        await self._background_task
        self._background_task = None

temporary_state async

temporary_state(
    *,
    system_prompts: list[AnyPromptType] | None = None,
    replace_prompts: bool = False,
    tools: list[ToolType] | None = None,
    replace_tools: bool = False,
    history: list[AnyPromptType] | SessionQuery | None = None,
    replace_history: bool = False,
    pause_routing: bool = False,
    model: ModelType | None = None,
    provider: AgentProvider | None = None,
) -> AsyncIterator[Self]

Temporarily modify agent state.

Parameters:

Name	Type	Description	Default
system_prompts	list[AnyPromptType] | None	Temporary system prompts to use	None
replace_prompts	bool	Whether to replace existing prompts	False
tools	list[ToolType] | None	Temporary tools to make available	None
replace_tools	bool	Whether to replace existing tools	False
history	list[AnyPromptType] | SessionQuery | None	Conversation history (prompts or query)	None
replace_history	bool	Whether to replace existing history	False
pause_routing	bool	Whether to pause message routing	False
model	ModelType | None	Temporary model override	None
provider	AgentProvider | None	Temporary provider override	None

Source code in src/llmling_agent/agent/agent.py

@asynccontextmanager
async def temporary_state(
    self,
    *,
    system_prompts: list[AnyPromptType] | None = None,
    replace_prompts: bool = False,
    tools: list[ToolType] | None = None,
    replace_tools: bool = False,
    history: list[AnyPromptType] | SessionQuery | None = None,
    replace_history: bool = False,
    pause_routing: bool = False,
    model: ModelType | None = None,
    provider: AgentProvider | None = None,
) -> AsyncIterator[Self]:
    """Temporarily modify agent state.

    Args:
        system_prompts: Temporary system prompts to use
        replace_prompts: Whether to replace existing prompts
        tools: Temporary tools to make available
        replace_tools: Whether to replace existing tools
        history: Conversation history (prompts or query)
        replace_history: Whether to replace existing history
        pause_routing: Whether to pause message routing
        model: Temporary model override
        provider: Temporary provider override
    """
    old_model = self._provider.model if hasattr(self._provider, "model") else None  # pyright: ignore
    old_provider = self._provider

    async with AsyncExitStack() as stack:
        # System prompts (async)
        if system_prompts is not None:
            await stack.enter_async_context(
                self.sys_prompts.temporary_prompt(
                    system_prompts, exclusive=replace_prompts
                )
            )

        # Tools (sync)
        if tools is not None:
            stack.enter_context(
                self.tools.temporary_tools(tools, exclusive=replace_tools)
            )

        # History (async)
        if history is not None:
            await stack.enter_async_context(
                self.conversation.temporary_state(
                    history, replace_history=replace_history
                )
            )

        # Routing (async)
        if pause_routing:
            await stack.enter_async_context(self.connections.paused_routing())

        # Model/Provider
        if provider is not None:
            self._provider = provider
        elif model is not None:
            self._provider.set_model(model)

        try:
            yield self
        finally:
            # Restore model/provider
            if provider is not None:
                self._provider = old_provider
            elif model is not None and old_model:
                self._provider.set_model(old_model)

to_structured

to_structured(
    result_type: None,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
) -> Self

to_structured(
    result_type: type[TResult] | str | ResponseDefinition,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
) -> StructuredAgent[TDeps, TResult]

to_structured(
    result_type: type[TResult] | str | ResponseDefinition | None,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
) -> StructuredAgent[TDeps, TResult] | Self

Convert this agent to a structured agent.

If result_type is None, returns self unchanged (no wrapping). Otherwise creates a StructuredAgent wrapper.

Parameters:

Name	Type	Description	Default
result_type	type[TResult] | str | ResponseDefinition | None	Type for structured responses. Can be: - A Python type (Pydantic model) - Name of response definition from context - Complete response definition - None to skip wrapping	required
tool_name	str | None	Optional override for result tool name	None
tool_description	str | None	Optional override for result tool description	None

Returns:

Type	Description
StructuredAgent[TDeps, TResult] | Self	Either StructuredAgent wrapper or self unchanged

from llmling_agent.agent import StructuredAgent

Source code in src/llmling_agent/agent/agent.py

def to_structured[TResult](
    self,
    result_type: type[TResult] | str | ResponseDefinition | None,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
) -> StructuredAgent[TDeps, TResult] | Self:
    """Convert this agent to a structured agent.

    If result_type is None, returns self unchanged (no wrapping).
    Otherwise creates a StructuredAgent wrapper.

    Args:
        result_type: Type for structured responses. Can be:
            - A Python type (Pydantic model)
            - Name of response definition from context
            - Complete response definition
            - None to skip wrapping
        tool_name: Optional override for result tool name
        tool_description: Optional override for result tool description

    Returns:
        Either StructuredAgent wrapper or self unchanged
    from llmling_agent.agent import StructuredAgent
    """
    if result_type is None:
        return self

    from llmling_agent.agent import StructuredAgent

    return StructuredAgent(
        self,
        result_type=result_type,
        tool_name=tool_name,
        tool_description=tool_description,
    )

to_tool

to_tool(
    *,
    name: str | None = None,
    reset_history_on_run: bool = True,
    pass_message_history: bool = False,
    share_context: bool = False,
    parent: AnyAgent[Any, Any] | None = None,
) -> Tool

Create a tool from this agent.

Parameters:

Name	Type	Description	Default
name	str | None	Optional tool name override	None
reset_history_on_run	bool	Clear agent's history before each run	True
pass_message_history	bool	Pass parent's message history to agent	False
share_context	bool	Whether to pass parent's context/deps	False
parent	AnyAgent[Any, Any] | None	Optional parent agent for history/context sharing	None

Source code in src/llmling_agent/agent/agent.py

def to_tool(
    self,
    *,
    name: str | None = None,
    reset_history_on_run: bool = True,
    pass_message_history: bool = False,
    share_context: bool = False,
    parent: AnyAgent[Any, Any] | None = None,
) -> Tool:
    """Create a tool from this agent.

    Args:
        name: Optional tool name override
        reset_history_on_run: Clear agent's history before each run
        pass_message_history: Pass parent's message history to agent
        share_context: Whether to pass parent's context/deps
        parent: Optional parent agent for history/context sharing
    """
    tool_name = f"ask_{self.name}"

    async def wrapped_tool(prompt: str) -> str:
        if pass_message_history and not parent:
            msg = "Parent agent required for message history sharing"
            raise ToolError(msg)

        if reset_history_on_run:
            self.conversation.clear()

        history = None
        if pass_message_history and parent:
            history = parent.conversation.get_history()
            old = self.conversation.get_history()
            self.conversation.set_history(history)
        result = await self.run(prompt, result_type=self._result_type)
        if history:
            self.conversation.set_history(old)
        return result.data

    normalized_name = self.name.replace("_", " ").title()
    docstring = f"Get expert answer from specialized agent: {normalized_name}"
    if self.description:
        docstring = f"{docstring}\n\n{self.description}"

    wrapped_tool.__doc__ = docstring
    wrapped_tool.__name__ = tool_name

    return Tool.from_callable(
        wrapped_tool,
        name_override=tool_name,
        description_override=docstring,
    )

wait async

wait() -> ChatMessage[TResult]

Wait for background execution to complete.

Source code in src/llmling_agent/agent/agent.py

async def wait(self) -> ChatMessage[TResult]:
    """Wait for background execution to complete."""
    if not self._background_task:
        msg = "No background task running"
        raise RuntimeError(msg)
    if self._infinite:
        msg = "Cannot wait on infinite execution"
        raise RuntimeError(msg)
    try:
        return await self._background_task
    finally:
        self._background_task = None

AgentConfig

Bases: NodeConfig

Configuration for a single agent in the system.

Defines an agent's complete configuration including its model, environment, capabilities, and behavior settings. Each agent can have its own: - Language model configuration - Environment setup (tools and resources) - Response type definitions - System prompts and default user prompts - Role-based capabilities

The configuration can be loaded from YAML or created programmatically.

Source code in src/llmling_agent/models/agents.py

class AgentConfig(NodeConfig):
    """Configuration for a single agent in the system.

    Defines an agent's complete configuration including its model, environment,
    capabilities, and behavior settings. Each agent can have its own:
    - Language model configuration
    - Environment setup (tools and resources)
    - Response type definitions
    - System prompts and default user prompts
    - Role-based capabilities

    The configuration can be loaded from YAML or created programmatically.
    """

    provider: ProviderConfig | ProviderName = "pydantic_ai"
    """Provider configuration or shorthand type"""

    inherits: str | None = None
    """Name of agent config to inherit from"""

    model: str | AnyModelConfig | None = None
    """The model to use for this agent. Can be either a simple model name
    string (e.g. 'openai:gpt-4') or a structured model definition."""

    tools: list[ToolConfig | str] = Field(default_factory=list)
    """A list of tools to register with this agent."""

    toolsets: list[ToolsetConfig] = Field(default_factory=list)
    """Toolset configurations for extensible tool collections."""

    environment: str | AgentEnvironment | None = None
    """Environments configuration (path or object)"""

    capabilities: Capabilities = Field(default_factory=Capabilities)
    """Current agent's capabilities."""

    session: str | SessionQuery | MemoryConfig | None = None
    """Session configuration for conversation recovery."""

    result_type: str | ResponseDefinition | None = None
    """Name of the response definition to use"""

    retries: int = 1
    """Number of retries for failed operations (maps to pydantic-ai's retries)"""

    result_tool_name: str = "final_result"
    """Name of the tool used for structured responses"""

    result_tool_description: str | None = None
    """Custom description for the result tool"""

    result_retries: int | None = None
    """Max retries for result validation"""

    end_strategy: EndStrategy = "early"
    """The strategy for handling multiple tool calls when a final result is found"""

    avatar: str | None = None
    """URL or path to agent's avatar image"""

    system_prompts: list[str] = Field(default_factory=list)
    """System prompts for the agent"""

    library_system_prompts: list[str] = Field(default_factory=list)
    """System prompts for the agent from the library"""

    user_prompts: list[str] = Field(default_factory=list)
    """Default user prompts for the agent"""

    # context_sources: list[ContextSource] = Field(default_factory=list)
    # """Initial context sources to load"""

    config_file_path: str | None = None
    """Config file path for resolving environment."""

    knowledge: Knowledge | None = None
    """Knowledge sources for this agent."""

    workers: list[WorkerConfig] = Field(default_factory=list)
    """Worker agents which will be available as tools."""

    requires_tool_confirmation: ToolConfirmationMode = "per_tool"
    """How to handle tool confirmation:
    - "always": Always require confirmation for all tools
    - "never": Never require confirmation (ignore tool settings)
    - "per_tool": Use individual tool settings
    """

    debug: bool = False
    """Enable debug output for this agent."""

    def is_structured(self) -> bool:
        """Check if this config defines a structured agent."""
        return self.result_type is not None

    @model_validator(mode="before")
    @classmethod
    def validate_result_type(cls, data: dict[str, Any]) -> dict[str, Any]:
        """Convert result type and apply its settings."""
        result_type = data.get("result_type")
        if isinstance(result_type, dict):
            # Extract response-specific settings
            tool_name = result_type.pop("result_tool_name", None)
            tool_description = result_type.pop("result_tool_description", None)
            retries = result_type.pop("result_retries", None)

            # Convert remaining dict to ResponseDefinition
            if "type" not in result_type:
                result_type["type"] = "inline"
            data["result_type"] = InlineResponseDefinition(**result_type)

            # Apply extracted settings to agent config
            if tool_name:
                data["result_tool_name"] = tool_name
            if tool_description:
                data["result_tool_description"] = tool_description
            if retries is not None:
                data["result_retries"] = retries

        return data

    @model_validator(mode="before")
    @classmethod
    def handle_model_types(cls, data: dict[str, Any]) -> dict[str, Any]:
        """Convert model inputs to appropriate format."""
        model = data.get("model")
        match model:
            case str():
                data["model"] = {"type": "string", "identifier": model}
        return data

    async def get_toolsets(self) -> list[ResourceProvider]:
        """Get all resource providers for this agent."""
        providers: list[ResourceProvider] = []

        # Add providers from toolsets
        for toolset_config in self.toolsets:
            try:
                provider = toolset_config.get_provider()
                providers.append(provider)
            except Exception as e:
                logger.exception(
                    "Failed to create provider for toolset: %r", toolset_config
                )
                msg = f"Failed to create provider for toolset: {e}"
                raise ValueError(msg) from e

        return providers

    def get_tool_provider(self) -> ResourceProvider | None:
        """Get tool provider for this agent."""
        from llmling_agent.tools.base import Tool

        # Create provider for static tools
        if not self.tools:
            return None
        static_tools: list[Tool] = []
        for tool_config in self.tools:
            try:
                match tool_config:
                    case str():
                        if tool_config.startswith("crewai_tools"):
                            obj = import_class(tool_config)()
                            static_tools.append(Tool.from_crewai_tool(obj))
                        elif tool_config.startswith("langchain"):
                            obj = import_class(tool_config)()
                            static_tools.append(Tool.from_langchain_tool(obj))
                        else:
                            tool = Tool.from_callable(tool_config)
                            static_tools.append(tool)
                    case BaseToolConfig():
                        static_tools.append(tool_config.get_tool())
            except Exception:
                logger.exception("Failed to load tool %r", tool_config)
                continue

        return StaticResourceProvider(name="builtin", tools=static_tools)

    def get_session_config(self) -> MemoryConfig:
        """Get resolved memory configuration."""
        match self.session:
            case str() | UUID():
                return MemoryConfig(session=SessionQuery(name=str(self.session)))
            case SessionQuery():
                return MemoryConfig(session=self.session)
            case MemoryConfig():
                return self.session
            case None:
                return MemoryConfig()

    def get_system_prompts(self) -> list[BasePrompt]:
        """Get all system prompts as BasePrompts."""
        prompts: list[BasePrompt] = []
        for prompt in self.system_prompts:
            match prompt:
                case str():
                    # Convert string to StaticPrompt
                    static_prompt = StaticPrompt(
                        name="system",
                        description="System prompt",
                        messages=[PromptMessage(role="system", content=prompt)],
                    )
                    prompts.append(static_prompt)
                case BasePrompt():
                    prompts.append(prompt)
        return prompts

    def get_provider(self) -> AgentProvider:
        """Get resolved provider instance.

        Creates provider instance based on configuration:
        - Full provider config: Use as-is
        - Shorthand type: Create default provider config
        """
        # If string shorthand is used, convert to default provider config
        from llmling_agent_config.providers import (
            CallbackProviderConfig,
            HumanProviderConfig,
            LiteLLMProviderConfig,
            PydanticAIProviderConfig,
        )

        provider_config = self.provider
        if isinstance(provider_config, str):
            match provider_config:
                case "pydantic_ai":
                    provider_config = PydanticAIProviderConfig(model=self.model)
                case "human":
                    provider_config = HumanProviderConfig()
                case "litellm":
                    provider_config = LiteLLMProviderConfig(
                        model=self.model if isinstance(self.model, str) else None
                    )
                case _:
                    try:
                        fn = import_callable(provider_config)
                        provider_config = CallbackProviderConfig(fn=fn)
                    except Exception:  # noqa: BLE001
                        msg = f"Invalid provider type: {provider_config}"
                        raise ValueError(msg)  # noqa: B904

        # Create provider instance from config
        return provider_config.get_provider()

    def render_system_prompts(self, context: dict[str, Any] | None = None) -> list[str]:
        """Render system prompts with context."""
        if not context:
            # Default context
            context = {"name": self.name, "id": 1, "model": self.model}
        return [render_prompt(p, {"agent": context}) for p in self.system_prompts]

    def get_config(self) -> Config:
        """Get configuration for this agent."""
        match self.environment:
            case None:
                # Create minimal config
                caps = LLMCapabilitiesConfig()
                global_settings = GlobalSettings(llm_capabilities=caps)
                return Config(global_settings=global_settings)
            case str() as path:
                # Backward compatibility: treat as file path
                resolved = self._resolve_environment_path(path, self.config_file_path)
                return Config.from_file(resolved)
            case FileEnvironment(uri=uri) as env:
                # Handle FileEnvironment instance
                resolved = env.get_file_path()
                return Config.from_file(resolved)
            case {"type": "file", "uri": uri}:
                # Handle raw dict matching file environment structure
                return Config.from_file(uri)
            case {"type": "inline", "config": config}:
                return config
            case InlineEnvironment() as config:
                return config
            case _:
                msg = f"Invalid environment configuration: {self.environment}"
                raise ValueError(msg)

    def get_environment_path(self) -> str | None:
        """Get environment file path if available."""
        match self.environment:
            case str() as path:
                return self._resolve_environment_path(path, self.config_file_path)
            case {"type": "file", "uri": uri} | FileEnvironment(uri=uri):
                return uri
            case _:
                return None

    @staticmethod
    def _resolve_environment_path(env: str, config_file_path: str | None = None) -> str:
        """Resolve environment path from config store or relative path."""
        from upath import UPath

        try:
            config_store = ConfigStore()
            return config_store.get_config(env)
        except KeyError:
            if config_file_path:
                base_dir = UPath(config_file_path).parent
                return str(base_dir / env)
            return env

avatar class-attribute instance-attribute

avatar: str | None = None

URL or path to agent's avatar image

capabilities class-attribute instance-attribute

capabilities: Capabilities = Field(default_factory=Capabilities)

Current agent's capabilities.

config_file_path class-attribute instance-attribute

config_file_path: str | None = None

Config file path for resolving environment.

debug class-attribute instance-attribute

debug: bool = False

Enable debug output for this agent.

end_strategy class-attribute instance-attribute

end_strategy: EndStrategy = 'early'

The strategy for handling multiple tool calls when a final result is found

environment class-attribute instance-attribute

environment: str | AgentEnvironment | None = None

Environments configuration (path or object)

inherits class-attribute instance-attribute

inherits: str | None = None

Name of agent config to inherit from

knowledge class-attribute instance-attribute

knowledge: Knowledge | None = None

Knowledge sources for this agent.

library_system_prompts class-attribute instance-attribute

library_system_prompts: list[str] = Field(default_factory=list)

System prompts for the agent from the library

model class-attribute instance-attribute

model: str | AnyModelConfig | None = None

The model to use for this agent. Can be either a simple model name string (e.g. 'openai:gpt-4') or a structured model definition.

provider class-attribute instance-attribute

provider: ProviderConfig | ProviderName = 'pydantic_ai'

Provider configuration or shorthand type

requires_tool_confirmation class-attribute instance-attribute

requires_tool_confirmation: ToolConfirmationMode = 'per_tool'

How to handle tool confirmation: - "always": Always require confirmation for all tools - "never": Never require confirmation (ignore tool settings) - "per_tool": Use individual tool settings

result_retries class-attribute instance-attribute

result_retries: int | None = None

Max retries for result validation

result_tool_description class-attribute instance-attribute

result_tool_description: str | None = None

Custom description for the result tool

result_tool_name class-attribute instance-attribute

result_tool_name: str = 'final_result'

Name of the tool used for structured responses

result_type class-attribute instance-attribute

result_type: str | ResponseDefinition | None = None

Name of the response definition to use

retries class-attribute instance-attribute

retries: int = 1

Number of retries for failed operations (maps to pydantic-ai's retries)

session class-attribute instance-attribute

session: str | SessionQuery | MemoryConfig | None = None

Session configuration for conversation recovery.

system_prompts class-attribute instance-attribute

system_prompts: list[str] = Field(default_factory=list)

System prompts for the agent

tools class-attribute instance-attribute

tools: list[ToolConfig | str] = Field(default_factory=list)

A list of tools to register with this agent.

toolsets class-attribute instance-attribute

toolsets: list[ToolsetConfig] = Field(default_factory=list)

Toolset configurations for extensible tool collections.

user_prompts class-attribute instance-attribute

user_prompts: list[str] = Field(default_factory=list)

Default user prompts for the agent

workers class-attribute instance-attribute

workers: list[WorkerConfig] = Field(default_factory=list)

Worker agents which will be available as tools.

get_config

get_config() -> Config

Get configuration for this agent.

Source code in src/llmling_agent/models/agents.py

def get_config(self) -> Config:
    """Get configuration for this agent."""
    match self.environment:
        case None:
            # Create minimal config
            caps = LLMCapabilitiesConfig()
            global_settings = GlobalSettings(llm_capabilities=caps)
            return Config(global_settings=global_settings)
        case str() as path:
            # Backward compatibility: treat as file path
            resolved = self._resolve_environment_path(path, self.config_file_path)
            return Config.from_file(resolved)
        case FileEnvironment(uri=uri) as env:
            # Handle FileEnvironment instance
            resolved = env.get_file_path()
            return Config.from_file(resolved)
        case {"type": "file", "uri": uri}:
            # Handle raw dict matching file environment structure
            return Config.from_file(uri)
        case {"type": "inline", "config": config}:
            return config
        case InlineEnvironment() as config:
            return config
        case _:
            msg = f"Invalid environment configuration: {self.environment}"
            raise ValueError(msg)

get_environment_path

get_environment_path() -> str | None

Get environment file path if available.

Source code in src/llmling_agent/models/agents.py

def get_environment_path(self) -> str | None:
    """Get environment file path if available."""
    match self.environment:
        case str() as path:
            return self._resolve_environment_path(path, self.config_file_path)
        case {"type": "file", "uri": uri} | FileEnvironment(uri=uri):
            return uri
        case _:
            return None

get_provider

get_provider() -> AgentProvider

Get resolved provider instance.

Creates provider instance based on configuration: - Full provider config: Use as-is - Shorthand type: Create default provider config

Source code in src/llmling_agent/models/agents.py

def get_provider(self) -> AgentProvider:
    """Get resolved provider instance.

    Creates provider instance based on configuration:
    - Full provider config: Use as-is
    - Shorthand type: Create default provider config
    """
    # If string shorthand is used, convert to default provider config
    from llmling_agent_config.providers import (
        CallbackProviderConfig,
        HumanProviderConfig,
        LiteLLMProviderConfig,
        PydanticAIProviderConfig,
    )

    provider_config = self.provider
    if isinstance(provider_config, str):
        match provider_config:
            case "pydantic_ai":
                provider_config = PydanticAIProviderConfig(model=self.model)
            case "human":
                provider_config = HumanProviderConfig()
            case "litellm":
                provider_config = LiteLLMProviderConfig(
                    model=self.model if isinstance(self.model, str) else None
                )
            case _:
                try:
                    fn = import_callable(provider_config)
                    provider_config = CallbackProviderConfig(fn=fn)
                except Exception:  # noqa: BLE001
                    msg = f"Invalid provider type: {provider_config}"
                    raise ValueError(msg)  # noqa: B904

    # Create provider instance from config
    return provider_config.get_provider()

get_session_config

get_session_config() -> MemoryConfig

Get resolved memory configuration.

Source code in src/llmling_agent/models/agents.py

def get_session_config(self) -> MemoryConfig:
    """Get resolved memory configuration."""
    match self.session:
        case str() | UUID():
            return MemoryConfig(session=SessionQuery(name=str(self.session)))
        case SessionQuery():
            return MemoryConfig(session=self.session)
        case MemoryConfig():
            return self.session
        case None:
            return MemoryConfig()

get_system_prompts

get_system_prompts() -> list[BasePrompt]

Get all system prompts as BasePrompts.

Source code in src/llmling_agent/models/agents.py

def get_system_prompts(self) -> list[BasePrompt]:
    """Get all system prompts as BasePrompts."""
    prompts: list[BasePrompt] = []
    for prompt in self.system_prompts:
        match prompt:
            case str():
                # Convert string to StaticPrompt
                static_prompt = StaticPrompt(
                    name="system",
                    description="System prompt",
                    messages=[PromptMessage(role="system", content=prompt)],
                )
                prompts.append(static_prompt)
            case BasePrompt():
                prompts.append(prompt)
    return prompts

get_tool_provider

get_tool_provider() -> ResourceProvider | None

Get tool provider for this agent.

Source code in src/llmling_agent/models/agents.py

def get_tool_provider(self) -> ResourceProvider | None:
    """Get tool provider for this agent."""
    from llmling_agent.tools.base import Tool

    # Create provider for static tools
    if not self.tools:
        return None
    static_tools: list[Tool] = []
    for tool_config in self.tools:
        try:
            match tool_config:
                case str():
                    if tool_config.startswith("crewai_tools"):
                        obj = import_class(tool_config)()
                        static_tools.append(Tool.from_crewai_tool(obj))
                    elif tool_config.startswith("langchain"):
                        obj = import_class(tool_config)()
                        static_tools.append(Tool.from_langchain_tool(obj))
                    else:
                        tool = Tool.from_callable(tool_config)
                        static_tools.append(tool)
                case BaseToolConfig():
                    static_tools.append(tool_config.get_tool())
        except Exception:
            logger.exception("Failed to load tool %r", tool_config)
            continue

    return StaticResourceProvider(name="builtin", tools=static_tools)

get_toolsets async

get_toolsets() -> list[ResourceProvider]

Get all resource providers for this agent.

Source code in src/llmling_agent/models/agents.py

async def get_toolsets(self) -> list[ResourceProvider]:
    """Get all resource providers for this agent."""
    providers: list[ResourceProvider] = []

    # Add providers from toolsets
    for toolset_config in self.toolsets:
        try:
            provider = toolset_config.get_provider()
            providers.append(provider)
        except Exception as e:
            logger.exception(
                "Failed to create provider for toolset: %r", toolset_config
            )
            msg = f"Failed to create provider for toolset: {e}"
            raise ValueError(msg) from e

    return providers

handle_model_types classmethod

handle_model_types(data: dict[str, Any]) -> dict[str, Any]

Convert model inputs to appropriate format.

Source code in src/llmling_agent/models/agents.py

@model_validator(mode="before")
@classmethod
def handle_model_types(cls, data: dict[str, Any]) -> dict[str, Any]:
    """Convert model inputs to appropriate format."""
    model = data.get("model")
    match model:
        case str():
            data["model"] = {"type": "string", "identifier": model}
    return data

is_structured

is_structured() -> bool

Check if this config defines a structured agent.

Source code in src/llmling_agent/models/agents.py

def is_structured(self) -> bool:
    """Check if this config defines a structured agent."""
    return self.result_type is not None

render_system_prompts

render_system_prompts(context: dict[str, Any] | None = None) -> list[str]

Render system prompts with context.

Source code in src/llmling_agent/models/agents.py

def render_system_prompts(self, context: dict[str, Any] | None = None) -> list[str]:
    """Render system prompts with context."""
    if not context:
        # Default context
        context = {"name": self.name, "id": 1, "model": self.model}
    return [render_prompt(p, {"agent": context}) for p in self.system_prompts]

validate_result_type classmethod

validate_result_type(data: dict[str, Any]) -> dict[str, Any]

Convert result type and apply its settings.

Source code in src/llmling_agent/models/agents.py

@model_validator(mode="before")
@classmethod
def validate_result_type(cls, data: dict[str, Any]) -> dict[str, Any]:
    """Convert result type and apply its settings."""
    result_type = data.get("result_type")
    if isinstance(result_type, dict):
        # Extract response-specific settings
        tool_name = result_type.pop("result_tool_name", None)
        tool_description = result_type.pop("result_tool_description", None)
        retries = result_type.pop("result_retries", None)

        # Convert remaining dict to ResponseDefinition
        if "type" not in result_type:
            result_type["type"] = "inline"
        data["result_type"] = InlineResponseDefinition(**result_type)

        # Apply extracted settings to agent config
        if tool_name:
            data["result_tool_name"] = tool_name
        if tool_description:
            data["result_tool_description"] = tool_description
        if retries is not None:
            data["result_retries"] = retries

    return data

AgentContext dataclass

Bases: NodeContext[TDeps]

Runtime context for agent execution.

Generically typed with AgentContext[Type of Dependencies]

Source code in src/llmling_agent/agent/context.py

@dataclass(kw_only=True)
class AgentContext[TDeps](NodeContext[TDeps]):
    """Runtime context for agent execution.

    Generically typed with AgentContext[Type of Dependencies]
    """

    capabilities: Capabilities
    """Current agent's capabilities."""

    config: AgentConfig
    """Current agent's specific configuration."""

    model_settings: dict[str, Any] = field(default_factory=dict)
    """Model-specific settings."""

    data: TDeps | None = None
    """Custom context data."""

    runtime: RuntimeConfig | None = None
    """Reference to the runtime configuration."""

    @classmethod
    def create_default(
        cls,
        name: str,
        capabilities: Capabilities | None = None,
        deps: TDeps | None = None,
        pool: AgentPool | None = None,
        input_provider: InputProvider | None = None,
    ) -> AgentContext[TDeps]:
        """Create a default agent context with minimal privileges.

        Args:
            name: Name of the agent
            capabilities: Optional custom capabilities (defaults to minimal access)
            deps: Optional dependencies for the agent
            pool: Optional pool the agent is part of
            input_provider: Optional input provider for the agent
        """
        from llmling_agent.config.capabilities import Capabilities
        from llmling_agent.models import AgentConfig, AgentsManifest

        caps = capabilities or Capabilities()
        defn = AgentsManifest()
        cfg = AgentConfig(name=name)
        return cls(
            input_provider=input_provider,
            node_name=name,
            capabilities=caps,
            definition=defn,
            config=cfg,
            data=deps,
            pool=pool,
        )

    @cached_property
    def converter(self) -> ConversionManager:
        """Get conversion manager from global config."""
        return ConversionManager(self.definition.conversion)

    # TODO: perhaps add agent directly to context?
    @property
    def agent(self) -> AnyAgent[TDeps, Any]:
        """Get the agent instance from the pool."""
        assert self.pool, "No agent pool available"
        assert self.node_name, "No agent name available"
        return self.pool.agents[self.node_name]

    async def handle_confirmation(
        self,
        tool: Tool,
        args: dict[str, Any],
    ) -> ConfirmationResult:
        """Handle tool execution confirmation.

        Returns True if:
        - No confirmation handler is set
        - Handler confirms the execution
        """
        provider = self.get_input_provider()
        mode = self.config.requires_tool_confirmation
        if (mode == "per_tool" and not tool.requires_confirmation) or mode == "never":
            return "allow"
        history = self.agent.conversation.get_history() if self.pool else []
        return await provider.get_tool_confirmation(self, tool, args, history)

agent property

agent: AnyAgent[TDeps, Any]

Get the agent instance from the pool.

capabilities instance-attribute

capabilities: Capabilities

Current agent's capabilities.

config instance-attribute

config: AgentConfig

Current agent's specific configuration.

converter cached property

converter: ConversionManager

Get conversion manager from global config.

data class-attribute instance-attribute

data: TDeps | None = None

Custom context data.

model_settings class-attribute instance-attribute

model_settings: dict[str, Any] = field(default_factory=dict)

Model-specific settings.

runtime class-attribute instance-attribute

runtime: RuntimeConfig | None = None

Reference to the runtime configuration.

create_default classmethod

create_default(
    name: str,
    capabilities: Capabilities | None = None,
    deps: TDeps | None = None,
    pool: AgentPool | None = None,
    input_provider: InputProvider | None = None,
) -> AgentContext[TDeps]

Create a default agent context with minimal privileges.

Parameters:

Name	Type	Description	Default
name	str	Name of the agent	required
capabilities	Capabilities | None	Optional custom capabilities (defaults to minimal access)	None
deps	TDeps | None	Optional dependencies for the agent	None
pool	AgentPool | None	Optional pool the agent is part of	None
input_provider	InputProvider | None	Optional input provider for the agent	None

Source code in src/llmling_agent/agent/context.py

@classmethod
def create_default(
    cls,
    name: str,
    capabilities: Capabilities | None = None,
    deps: TDeps | None = None,
    pool: AgentPool | None = None,
    input_provider: InputProvider | None = None,
) -> AgentContext[TDeps]:
    """Create a default agent context with minimal privileges.

    Args:
        name: Name of the agent
        capabilities: Optional custom capabilities (defaults to minimal access)
        deps: Optional dependencies for the agent
        pool: Optional pool the agent is part of
        input_provider: Optional input provider for the agent
    """
    from llmling_agent.config.capabilities import Capabilities
    from llmling_agent.models import AgentConfig, AgentsManifest

    caps = capabilities or Capabilities()
    defn = AgentsManifest()
    cfg = AgentConfig(name=name)
    return cls(
        input_provider=input_provider,
        node_name=name,
        capabilities=caps,
        definition=defn,
        config=cfg,
        data=deps,
        pool=pool,
    )

handle_confirmation async

handle_confirmation(tool: Tool, args: dict[str, Any]) -> ConfirmationResult

Handle tool execution confirmation.

Returns True if: - No confirmation handler is set - Handler confirms the execution

Source code in src/llmling_agent/agent/context.py

async def handle_confirmation(
    self,
    tool: Tool,
    args: dict[str, Any],
) -> ConfirmationResult:
    """Handle tool execution confirmation.

    Returns True if:
    - No confirmation handler is set
    - Handler confirms the execution
    """
    provider = self.get_input_provider()
    mode = self.config.requires_tool_confirmation
    if (mode == "per_tool" and not tool.requires_confirmation) or mode == "never":
        return "allow"
    history = self.agent.conversation.get_history() if self.pool else []
    return await provider.get_tool_confirmation(self, tool, args, history)

AgentPool

Bases: BaseRegistry[NodeName, MessageEmitter[Any, Any]]

Pool managing message processing nodes (agents and teams).

Acts as a unified registry for all nodes, providing: - Centralized node management and lookup - Shared dependency injection - Connection management - Resource coordination

Nodes can be accessed through: - nodes: All registered nodes (agents and teams) - agents: Only Agent instances - teams: Only Team instances

Source code in src/llmling_agent/delegation/pool.py

class AgentPool[TPoolDeps](BaseRegistry[NodeName, MessageEmitter[Any, Any]]):
    """Pool managing message processing nodes (agents and teams).

    Acts as a unified registry for all nodes, providing:
    - Centralized node management and lookup
    - Shared dependency injection
    - Connection management
    - Resource coordination

    Nodes can be accessed through:
    - nodes: All registered nodes (agents and teams)
    - agents: Only Agent instances
    - teams: Only Team instances
    """

    def __init__(
        self,
        manifest: StrPath | AgentsManifest | None = None,
        *,
        shared_deps: TPoolDeps | None = None,
        connect_nodes: bool = True,
        input_provider: InputProvider | None = None,
        parallel_load: bool = True,
    ):
        """Initialize agent pool with immediate agent creation.

        Args:
            manifest: Agent configuration manifest
            shared_deps: Dependencies to share across all nodes
            connect_nodes: Whether to set up forwarding connections
            input_provider: Input provider for tool / step confirmations / HumanAgents
            parallel_load: Whether to load nodes in parallel (async)

        Raises:
            ValueError: If manifest contains invalid node configurations
            RuntimeError: If node initialization fails
        """
        super().__init__()
        from llmling_agent.mcp_server.manager import MCPManager
        from llmling_agent.models.manifest import AgentsManifest
        from llmling_agent.storage import StorageManager

        match manifest:
            case None:
                self.manifest = AgentsManifest()
            case str() | os.PathLike():
                self.manifest = AgentsManifest.from_file(manifest)
            case AgentsManifest():
                self.manifest = manifest
            case _:
                msg = f"Invalid config path: {manifest}"
                raise ValueError(msg)
        self.shared_deps = shared_deps
        self._input_provider = input_provider
        self.exit_stack = AsyncExitStack()
        self.parallel_load = parallel_load
        self.storage = StorageManager(self.manifest.storage)
        self.connection_registry = ConnectionRegistry()
        self.mcp = MCPManager(
            name="pool_mcp", servers=self.manifest.get_mcp_servers(), owner="pool"
        )
        self._tasks = TaskRegistry()
        # Register tasks from manifest
        for name, task in self.manifest.jobs.items():
            self._tasks.register(name, task)
        self.pool_talk = TeamTalk[Any].from_nodes(list(self.nodes.values()))
        if self.manifest.pool_server and self.manifest.pool_server.enabled:
            from llmling_agent.resource_providers.pool import PoolResourceProvider
            from llmling_agent_mcp.server import LLMLingServer

            provider = PoolResourceProvider(
                self, zed_mode=self.manifest.pool_server.zed_mode
            )
            self.server: LLMLingServer | None = LLMLingServer(
                provider=provider,
                config=self.manifest.pool_server,
            )
        else:
            self.server = None
        # Create requested agents immediately
        for name in self.manifest.agents:
            agent = self.manifest.get_agent(name, deps=shared_deps)
            self.register(name, agent)

        # Then set up worker relationships
        for agent in self.agents.values():
            self.setup_agent_workers(agent)
        self._create_teams()
        # Set up forwarding connections
        if connect_nodes:
            self._connect_nodes()

    async def __aenter__(self) -> Self:
        """Enter async context and initialize all agents."""
        try:
            # Add MCP tool provider to all agents
            agents = list(self.agents.values())
            teams = list(self.teams.values())
            for agent in agents:
                agent.tools.add_provider(self.mcp)

            # Collect all components to initialize
            components: list[AbstractAsyncContextManager[Any]] = [
                self.mcp,
                *agents,
                *teams,
            ]

            # Add MCP server if configured
            if self.server:
                components.append(self.server)

            # Initialize all components
            if self.parallel_load:
                await asyncio.gather(
                    *(self.exit_stack.enter_async_context(c) for c in components)
                )
            else:
                for component in components:
                    await self.exit_stack.enter_async_context(component)

        except Exception as e:
            await self.cleanup()
            msg = "Failed to initialize agent pool"
            logger.exception(msg, exc_info=e)
            raise RuntimeError(msg) from e
        return self

    async def __aexit__(
        self,
        exc_type: type[BaseException] | None,
        exc_val: BaseException | None,
        exc_tb: TracebackType | None,
    ):
        """Exit async context."""
        # Remove MCP tool provider from all agents
        for agent in self.agents.values():
            if self.mcp in agent.tools.providers:
                agent.tools.remove_provider(self.mcp)
        await self.cleanup()

    async def cleanup(self):
        """Clean up all agents."""
        await self.exit_stack.aclose()
        self.clear()

    @overload
    def create_team_run(
        self,
        agents: Sequence[str],
        validator: MessageNode[Any, TResult] | None = None,
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> TeamRun[TPoolDeps, TResult]: ...

    @overload
    def create_team_run[TDeps, TResult](
        self,
        agents: Sequence[MessageNode[TDeps, Any]],
        validator: MessageNode[Any, TResult] | None = None,
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> TeamRun[TDeps, TResult]: ...

    @overload
    def create_team_run(
        self,
        agents: Sequence[AgentName | MessageNode[Any, Any]],
        validator: MessageNode[Any, TResult] | None = None,
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> TeamRun[Any, TResult]: ...

    def create_team_run(
        self,
        agents: Sequence[AgentName | MessageNode[Any, Any]] | None = None,
        validator: MessageNode[Any, TResult] | None = None,
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> TeamRun[Any, TResult]:
        """Create a a sequential TeamRun from a list of Agents.

        Args:
            agents: List of agent names or team/agent instances (all if None)
            validator: Node to validate the results of the TeamRun
            name: Optional name for the team
            description: Optional description for the team
            shared_prompt: Optional prompt for all agents
            picker: Agent to use for picking agents
            num_picks: Number of agents to pick
            pick_prompt: Prompt to use for picking agents
        """
        from llmling_agent.delegation.teamrun import TeamRun

        if agents is None:
            agents = list(self.agents.keys())

        # First resolve/configure agents
        resolved_agents: list[MessageNode[Any, Any]] = []
        for agent in agents:
            if isinstance(agent, str):
                agent = self.get_agent(agent)
            resolved_agents.append(agent)
        team = TeamRun(
            resolved_agents,
            name=name,
            description=description,
            validator=validator,
            shared_prompt=shared_prompt,
            picker=picker,
            num_picks=num_picks,
            pick_prompt=pick_prompt,
        )
        if name:
            self[name] = team
        return team

    @overload
    def create_team(self, agents: Sequence[str]) -> Team[TPoolDeps]: ...

    @overload
    def create_team[TDeps](
        self,
        agents: Sequence[MessageNode[TDeps, Any]],
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> Team[TDeps]: ...

    @overload
    def create_team(
        self,
        agents: Sequence[AgentName | MessageNode[Any, Any]],
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> Team[Any]: ...

    def create_team(
        self,
        agents: Sequence[AgentName | MessageNode[Any, Any]] | None = None,
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ) -> Team[Any]:
        """Create a group from agent names or instances.

        Args:
            agents: List of agent names or instances (all if None)
            name: Optional name for the team
            description: Optional description for the team
            shared_prompt: Optional prompt for all agents
            picker: Agent to use for picking agents
            num_picks: Number of agents to pick
            pick_prompt: Prompt to use for picking agents
        """
        from llmling_agent.delegation.team import Team

        if agents is None:
            agents = list(self.agents.keys())

        # First resolve/configure agents
        resolved_agents: list[MessageNode[Any, Any]] = []
        for agent in agents:
            if isinstance(agent, str):
                agent = self.get_agent(agent)
            resolved_agents.append(agent)

        team = Team(
            name=name,
            description=description,
            agents=resolved_agents,
            shared_prompt=shared_prompt,
            picker=picker,
            num_picks=num_picks,
            pick_prompt=pick_prompt,
        )
        if name:
            self[name] = team
        return team

    @asynccontextmanager
    async def track_message_flow(self) -> AsyncIterator[MessageFlowTracker]:
        """Track message flow during a context."""
        tracker = MessageFlowTracker()
        self.connection_registry.message_flow.connect(tracker.track)
        try:
            yield tracker
        finally:
            self.connection_registry.message_flow.disconnect(tracker.track)

    async def run_event_loop(self):
        """Run pool in event-watching mode until interrupted."""
        import sys

        print("Starting event watch mode...")
        print("Active nodes: ", ", ".join(self.list_nodes()))
        print("Press Ctrl+C to stop")

        stop_event = asyncio.Event()

        if sys.platform != "win32":
            # Unix: Use signal handlers
            loop = asyncio.get_running_loop()
            for sig in (signal.SIGINT, signal.SIGTERM):
                loop.add_signal_handler(sig, stop_event.set)
            while True:
                await asyncio.sleep(1)
        else:
            # Windows: Use keyboard interrupt
            with suppress(KeyboardInterrupt):
                while True:
                    await asyncio.sleep(1)

    @property
    def agents(self) -> dict[str, AnyAgent[Any, Any]]:
        """Get agents dict (backward compatibility)."""
        return {
            i.name: i
            for i in self._items.values()
            if isinstance(i, Agent | StructuredAgent)
        }

    @property
    def teams(self) -> dict[str, BaseTeam[Any, Any]]:
        """Get agents dict (backward compatibility)."""
        from llmling_agent.delegation.base_team import BaseTeam

        return {i.name: i for i in self._items.values() if isinstance(i, BaseTeam)}

    @property
    def nodes(self) -> dict[str, MessageNode[Any, Any]]:
        """Get agents dict (backward compatibility)."""
        from llmling_agent import MessageNode

        return {i.name: i for i in self._items.values() if isinstance(i, MessageNode)}

    @property
    def event_nodes(self) -> dict[str, EventNode[Any]]:
        """Get agents dict (backward compatibility)."""
        from llmling_agent.messaging.eventnode import EventNode

        return {i.name: i for i in self._items.values() if isinstance(i, EventNode)}

    @property
    def node_events(self) -> DictEvents:
        """Get node events."""
        return self._items.events

    @property
    def _error_class(self) -> type[LLMLingError]:
        """Error class for agent operations."""
        return LLMLingError

    def _validate_item(
        self, item: MessageEmitter[Any, Any] | Any
    ) -> MessageEmitter[Any, Any]:
        """Validate and convert items before registration.

        Args:
            item: Item to validate

        Returns:
            Validated Node

        Raises:
            LLMlingError: If item is not a valid node
        """
        if not isinstance(item, MessageEmitter):
            msg = f"Item must be Agent or Team, got {type(item)}"
            raise self._error_class(msg)
        item.context.pool = self
        return item

    def _create_teams(self):
        """Create all teams in two phases to allow nesting."""
        # Phase 1: Create empty teams

        empty_teams: dict[str, BaseTeam[Any, Any]] = {}
        for name, config in self.manifest.teams.items():
            if config.mode == "parallel":
                empty_teams[name] = Team(
                    [], name=name, shared_prompt=config.shared_prompt
                )
            else:
                empty_teams[name] = TeamRun(
                    [], name=name, shared_prompt=config.shared_prompt
                )

        # Phase 2: Resolve members
        for name, config in self.manifest.teams.items():
            team = empty_teams[name]
            members: list[MessageNode[Any, Any]] = []
            for member in config.members:
                if member in self.agents:
                    members.append(self.agents[member])
                elif member in empty_teams:
                    members.append(empty_teams[member])
                else:
                    msg = f"Unknown team member: {member}"
                    raise ValueError(msg)
            team.agents.extend(members)
            self[name] = team

    def _connect_nodes(self):
        """Set up connections defined in manifest."""
        # Merge agent and team configs into one dict of nodes with connections
        for name, config in self.manifest.nodes.items():
            source = self[name]
            for target in config.connections or []:
                match target:
                    case NodeConnectionConfig():
                        if target.name not in self:
                            msg = f"Forward target {target.name} not found for {name}"
                            raise ValueError(msg)
                        target_node = self[target.name]
                    case FileConnectionConfig() | CallableConnectionConfig():
                        target_node = Agent(provider=target.get_provider())
                    case _:
                        msg = f"Invalid connection config: {target}"
                        raise ValueError(msg)

                source.connect_to(
                    target_node,  # type: ignore  # recognized as "Any | BaseTeam[Any, Any]" by mypy?
                    connection_type=target.connection_type,
                    name=name,
                    priority=target.priority,
                    delay=target.delay,
                    queued=target.queued,
                    queue_strategy=target.queue_strategy,
                    transform=target.transform,
                    filter_condition=target.filter_condition.check
                    if target.filter_condition
                    else None,
                    stop_condition=target.stop_condition.check
                    if target.stop_condition
                    else None,
                    exit_condition=target.exit_condition.check
                    if target.exit_condition
                    else None,
                )
                source.connections.set_wait_state(
                    target_node,
                    wait=target.wait_for_completion,
                )

    @overload
    async def clone_agent[TDeps](
        self,
        agent: AgentName | Agent[TDeps],
        new_name: AgentName | None = None,
        *,
        system_prompts: list[str] | None = None,
        template_context: dict[str, Any] | None = None,
    ) -> Agent[TDeps]: ...

    @overload
    async def clone_agent[TDeps, TResult](
        self,
        agent: StructuredAgent[TDeps, TResult],
        new_name: AgentName | None = None,
        *,
        system_prompts: list[str] | None = None,
        template_context: dict[str, Any] | None = None,
    ) -> StructuredAgent[TDeps, TResult]: ...

    async def clone_agent[TDeps, TAgentResult](
        self,
        agent: AgentName | AnyAgent[TDeps, TAgentResult],
        new_name: AgentName | None = None,
        *,
        system_prompts: list[str] | None = None,
        template_context: dict[str, Any] | None = None,
    ) -> AnyAgent[TDeps, TAgentResult]:
        """Create a copy of an agent.

        Args:
            agent: Agent instance or name to clone
            new_name: Optional name for the clone
            system_prompts: Optional different prompts
            template_context: Variables for template rendering

        Returns:
            The new agent instance
        """
        from llmling_agent.agent import Agent, StructuredAgent

        # Get original config
        if isinstance(agent, str):
            if agent not in self.manifest.agents:
                msg = f"Agent {agent} not found"
                raise KeyError(msg)
            config = self.manifest.agents[agent]
            original_agent: AnyAgent[Any, Any] = self.get_agent(agent)
        else:
            config = agent.context.config  # type: ignore
            original_agent = agent

        # Create new config
        new_config = config.model_copy(deep=True)

        # Apply overrides
        if system_prompts:
            new_config.system_prompts = system_prompts

        # Handle template rendering
        if template_context:
            new_config.system_prompts = new_config.render_system_prompts(template_context)

        # Create new agent with same runtime
        new_agent = Agent[TDeps](
            runtime=original_agent.runtime,
            context=original_agent.context,
            # result_type=original_agent.actual_type,
            provider=new_config.get_provider(),
            system_prompt=new_config.system_prompts,
            name=new_name or f"{config.name}_copy_{len(self.agents)}",
        )
        if isinstance(original_agent, StructuredAgent):
            new_agent = new_agent.to_structured(original_agent.actual_type)

        # Register in pool
        agent_name = new_agent.name
        self.manifest.agents[agent_name] = new_config
        self.register(agent_name, new_agent)
        return await self.exit_stack.enter_async_context(new_agent)

    @overload
    async def create_agent(
        self,
        name: AgentName,
        *,
        session: SessionIdType | SessionQuery = None,
        name_override: str | None = None,
    ) -> Agent[TPoolDeps]: ...

    @overload
    async def create_agent[TCustomDeps](
        self,
        name: AgentName,
        *,
        deps: TCustomDeps,
        session: SessionIdType | SessionQuery = None,
        name_override: str | None = None,
    ) -> Agent[TCustomDeps]: ...

    @overload
    async def create_agent[TResult](
        self,
        name: AgentName,
        *,
        return_type: type[TResult],
        session: SessionIdType | SessionQuery = None,
        name_override: str | None = None,
    ) -> StructuredAgent[TPoolDeps, TResult]: ...

    @overload
    async def create_agent[TCustomDeps, TResult](
        self,
        name: AgentName,
        *,
        deps: TCustomDeps,
        return_type: type[TResult],
        session: SessionIdType | SessionQuery = None,
        name_override: str | None = None,
    ) -> StructuredAgent[TCustomDeps, TResult]: ...

    async def create_agent(
        self,
        name: AgentName,
        *,
        deps: Any | None = None,
        return_type: Any | None = None,
        session: SessionIdType | SessionQuery = None,
        name_override: str | None = None,
    ) -> AnyAgent[Any, Any]:
        """Create a new agent instance from configuration.

        Args:
            name: Name of the agent configuration to use
            deps: Optional custom dependencies (overrides pool deps)
            return_type: Optional type for structured responses
            session: Optional session ID or query to recover conversation
            name_override: Optional different name for this instance

        Returns:
            New agent instance with the specified configuration

        Raises:
            KeyError: If agent configuration not found
            ValueError: If configuration is invalid
        """
        if name not in self.manifest.agents:
            msg = f"Agent configuration {name!r} not found"
            raise KeyError(msg)

        # Use Manifest.get_agent for proper initialization
        final_deps = deps if deps is not None else self.shared_deps
        agent = self.manifest.get_agent(name, deps=final_deps)
        # Override name if requested
        if name_override:
            agent.name = name_override

        # Set pool reference
        agent.context.pool = self

        # Handle session if provided
        if session:
            agent.conversation.load_history_from_database(session=session)

        # Initialize agent through exit stack
        agent = await self.exit_stack.enter_async_context(agent)

        # Override structured configuration if provided
        if return_type is not None:
            return agent.to_structured(return_type)

        return agent

    def setup_agent_workers(self, agent: AnyAgent[Any, Any]):
        """Set up workers for an agent from configuration."""
        for worker_config in agent.context.config.workers:
            try:
                worker = self.nodes[worker_config.name]
                match worker_config:
                    case TeamWorkerConfig():
                        agent.register_worker(worker)
                    case AgentWorkerConfig():
                        agent.register_worker(
                            worker,
                            name=worker_config.name,
                            reset_history_on_run=worker_config.reset_history_on_run,
                            pass_message_history=worker_config.pass_message_history,
                            share_context=worker_config.share_context,
                        )
            except KeyError as e:
                msg = f"Worker agent {worker_config.name!r} not found"
                raise ValueError(msg) from e

    @overload
    def get_agent(
        self,
        agent: AgentName | Agent[Any],
        *,
        model_override: str | None = None,
        session: SessionIdType | SessionQuery = None,
    ) -> Agent[TPoolDeps]: ...

    @overload
    def get_agent[TResult](
        self,
        agent: AgentName | Agent[Any],
        *,
        return_type: type[TResult],
        model_override: str | None = None,
        session: SessionIdType | SessionQuery = None,
    ) -> StructuredAgent[TPoolDeps, TResult]: ...

    @overload
    def get_agent[TCustomDeps](
        self,
        agent: AgentName | Agent[Any],
        *,
        deps: TCustomDeps,
        model_override: str | None = None,
        session: SessionIdType | SessionQuery = None,
    ) -> Agent[TCustomDeps]: ...

    @overload
    def get_agent[TCustomDeps, TResult](
        self,
        agent: AgentName | Agent[Any],
        *,
        deps: TCustomDeps,
        return_type: type[TResult],
        model_override: str | None = None,
        session: SessionIdType | SessionQuery = None,
    ) -> StructuredAgent[TCustomDeps, TResult]: ...

    def get_agent(
        self,
        agent: AgentName | Agent[Any],
        *,
        deps: Any | None = None,
        return_type: Any | None = None,
        model_override: str | None = None,
        session: SessionIdType | SessionQuery = None,
    ) -> AnyAgent[Any, Any]:
        """Get or configure an agent from the pool.

        This method provides flexible agent configuration with dependency injection:
        - Without deps: Agent uses pool's shared dependencies
        - With deps: Agent uses provided custom dependencies
        - With return_type: Returns a StructuredAgent with type validation

        Args:
            agent: Either agent name or instance
            deps: Optional custom dependencies (overrides shared deps)
            return_type: Optional type for structured responses
            model_override: Optional model override
            session: Optional session ID or query to recover conversation

        Returns:
            Either:
            - Agent[TPoolDeps] when using pool's shared deps
            - Agent[TCustomDeps] when custom deps provided
            - StructuredAgent when return_type provided

        Raises:
            KeyError: If agent name not found
            ValueError: If configuration is invalid
        """
        from llmling_agent.agent import Agent
        from llmling_agent.agent.context import AgentContext

        # Get base agent
        base = agent if isinstance(agent, Agent) else self.agents[agent]

        # Setup context and dependencies
        if base.context is None:
            base.context = AgentContext[Any].create_default(base.name)

        # Use custom deps if provided, otherwise use shared deps
        base.context.data = deps if deps is not None else self.shared_deps
        base.context.pool = self

        # Apply overrides
        if model_override:
            base.set_model(model_override)

        if session:
            base.conversation.load_history_from_database(session=session)

        # Convert to structured if needed
        if return_type is not None:
            return base.to_structured(return_type)

        return base

    def list_nodes(self) -> list[str]:
        """List available agent names."""
        return list(self.list_items())

    def get_job(self, name: str) -> Job[Any, Any]:
        return self._tasks[name]

    def register_task(self, name: str, task: Job[Any, Any]):
        self._tasks.register(name, task)

    @overload
    async def add_agent(
        self,
        name: AgentName,
        *,
        result_type: None = None,
        **kwargs: Unpack[AgentKwargs],
    ) -> Agent[Any]: ...

    @overload
    async def add_agent[TResult](
        self,
        name: AgentName,
        *,
        result_type: type[TResult] | str | ResponseDefinition,
        **kwargs: Unpack[AgentKwargs],
    ) -> StructuredAgent[Any, TResult]: ...

    async def add_agent(
        self,
        name: AgentName,
        *,
        result_type: type[Any] | str | ResponseDefinition | None = None,
        **kwargs: Unpack[AgentKwargs],
    ) -> Agent[Any] | StructuredAgent[Any, Any]:
        """Add a new permanent agent to the pool.

        Args:
            name: Name for the new agent
            result_type: Optional type for structured responses:
                - None: Regular unstructured agent
                - type: Python type for validation
                - str: Name of response definition
                - ResponseDefinition: Complete response definition
            **kwargs: Additional agent configuration

        Returns:
            Either a regular Agent or StructuredAgent depending on result_type
        """
        from llmling_agent.agent import Agent

        agent: AnyAgent[Any, Any] = Agent(name=name, **kwargs)
        agent.tools.add_provider(self.mcp)
        agent = await self.exit_stack.enter_async_context(agent)
        # Convert to structured if needed
        if result_type is not None:
            agent = agent.to_structured(result_type)
        self.register(name, agent)
        return agent

    def get_mermaid_diagram(
        self,
        include_details: bool = True,
    ) -> str:
        """Generate mermaid flowchart of all agents and their connections.

        Args:
            include_details: Whether to show connection details (types, queues, etc)
        """
        lines = ["flowchart LR"]

        # Add all agents as nodes
        for name in self.agents:
            lines.append(f"    {name}[{name}]")  # noqa: PERF401

        # Add all connections as edges
        for agent in self.agents.values():
            connections = agent.connections.get_connections()
            for talk in connections:
                talk = cast(Talk[Any], talk)  # help mypy understand it's a Talk
                source = talk.source.name
                for target in talk.targets:
                    if include_details:
                        details: list[str] = []
                        details.append(talk.connection_type)
                        if talk.queued:
                            details.append(f"queued({talk.queue_strategy})")
                        if fn := talk.filter_condition:  # type: ignore
                            details.append(f"filter:{fn.__name__}")
                        if fn := talk.stop_condition:  # type: ignore
                            details.append(f"stop:{fn.__name__}")
                        if fn := talk.exit_condition:  # type: ignore
                            details.append(f"exit:{fn.__name__}")

                        label = f"|{' '.join(details)}|" if details else ""
                        lines.append(f"    {source}--{label}-->{target.name}")
                    else:
                        lines.append(f"    {source}-->{target.name}")

        return "\n".join(lines)

agents property

agents: dict[str, AnyAgent[Any, Any]]

Get agents dict (backward compatibility).

event_nodes property

event_nodes: dict[str, EventNode[Any]]

Get agents dict (backward compatibility).

node_events property

node_events: DictEvents

Get node events.

nodes property

nodes: dict[str, MessageNode[Any, Any]]

Get agents dict (backward compatibility).

teams property

teams: dict[str, BaseTeam[Any, Any]]

Get agents dict (backward compatibility).

__aenter__ async

__aenter__() -> Self

Enter async context and initialize all agents.

Source code in src/llmling_agent/delegation/pool.py

async def __aenter__(self) -> Self:
    """Enter async context and initialize all agents."""
    try:
        # Add MCP tool provider to all agents
        agents = list(self.agents.values())
        teams = list(self.teams.values())
        for agent in agents:
            agent.tools.add_provider(self.mcp)

        # Collect all components to initialize
        components: list[AbstractAsyncContextManager[Any]] = [
            self.mcp,
            *agents,
            *teams,
        ]

        # Add MCP server if configured
        if self.server:
            components.append(self.server)

        # Initialize all components
        if self.parallel_load:
            await asyncio.gather(
                *(self.exit_stack.enter_async_context(c) for c in components)
            )
        else:
            for component in components:
                await self.exit_stack.enter_async_context(component)

    except Exception as e:
        await self.cleanup()
        msg = "Failed to initialize agent pool"
        logger.exception(msg, exc_info=e)
        raise RuntimeError(msg) from e
    return self

__aexit__ async

__aexit__(
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
)

Exit async context.

Source code in src/llmling_agent/delegation/pool.py

async def __aexit__(
    self,
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
):
    """Exit async context."""
    # Remove MCP tool provider from all agents
    for agent in self.agents.values():
        if self.mcp in agent.tools.providers:
            agent.tools.remove_provider(self.mcp)
    await self.cleanup()

__init__

__init__(
    manifest: StrPath | AgentsManifest | None = None,
    *,
    shared_deps: TPoolDeps | None = None,
    connect_nodes: bool = True,
    input_provider: InputProvider | None = None,
    parallel_load: bool = True,
)

Initialize agent pool with immediate agent creation.

Parameters:

Name	Type	Description	Default
manifest	StrPath | AgentsManifest | None	Agent configuration manifest	None
shared_deps	TPoolDeps | None	Dependencies to share across all nodes	None
connect_nodes	bool	Whether to set up forwarding connections	True
input_provider	InputProvider | None	Input provider for tool / step confirmations / HumanAgents	None
parallel_load	bool	Whether to load nodes in parallel (async)	True

Raises:

Type	Description
ValueError	If manifest contains invalid node configurations
RuntimeError	If node initialization fails

Source code in src/llmling_agent/delegation/pool.py

def __init__(
    self,
    manifest: StrPath | AgentsManifest | None = None,
    *,
    shared_deps: TPoolDeps | None = None,
    connect_nodes: bool = True,
    input_provider: InputProvider | None = None,
    parallel_load: bool = True,
):
    """Initialize agent pool with immediate agent creation.

    Args:
        manifest: Agent configuration manifest
        shared_deps: Dependencies to share across all nodes
        connect_nodes: Whether to set up forwarding connections
        input_provider: Input provider for tool / step confirmations / HumanAgents
        parallel_load: Whether to load nodes in parallel (async)

    Raises:
        ValueError: If manifest contains invalid node configurations
        RuntimeError: If node initialization fails
    """
    super().__init__()
    from llmling_agent.mcp_server.manager import MCPManager
    from llmling_agent.models.manifest import AgentsManifest
    from llmling_agent.storage import StorageManager

    match manifest:
        case None:
            self.manifest = AgentsManifest()
        case str() | os.PathLike():
            self.manifest = AgentsManifest.from_file(manifest)
        case AgentsManifest():
            self.manifest = manifest
        case _:
            msg = f"Invalid config path: {manifest}"
            raise ValueError(msg)
    self.shared_deps = shared_deps
    self._input_provider = input_provider
    self.exit_stack = AsyncExitStack()
    self.parallel_load = parallel_load
    self.storage = StorageManager(self.manifest.storage)
    self.connection_registry = ConnectionRegistry()
    self.mcp = MCPManager(
        name="pool_mcp", servers=self.manifest.get_mcp_servers(), owner="pool"
    )
    self._tasks = TaskRegistry()
    # Register tasks from manifest
    for name, task in self.manifest.jobs.items():
        self._tasks.register(name, task)
    self.pool_talk = TeamTalk[Any].from_nodes(list(self.nodes.values()))
    if self.manifest.pool_server and self.manifest.pool_server.enabled:
        from llmling_agent.resource_providers.pool import PoolResourceProvider
        from llmling_agent_mcp.server import LLMLingServer

        provider = PoolResourceProvider(
            self, zed_mode=self.manifest.pool_server.zed_mode
        )
        self.server: LLMLingServer | None = LLMLingServer(
            provider=provider,
            config=self.manifest.pool_server,
        )
    else:
        self.server = None
    # Create requested agents immediately
    for name in self.manifest.agents:
        agent = self.manifest.get_agent(name, deps=shared_deps)
        self.register(name, agent)

    # Then set up worker relationships
    for agent in self.agents.values():
        self.setup_agent_workers(agent)
    self._create_teams()
    # Set up forwarding connections
    if connect_nodes:
        self._connect_nodes()

add_agent async

add_agent(
    name: AgentName, *, result_type: None = None, **kwargs: Unpack[AgentKwargs]
) -> Agent[Any]

add_agent(
    name: AgentName,
    *,
    result_type: type[TResult] | str | ResponseDefinition,
    **kwargs: Unpack[AgentKwargs],
) -> StructuredAgent[Any, TResult]

add_agent(
    name: AgentName,
    *,
    result_type: type[Any] | str | ResponseDefinition | None = None,
    **kwargs: Unpack[AgentKwargs],
) -> Agent[Any] | StructuredAgent[Any, Any]

Add a new permanent agent to the pool.

Parameters:

Name	Type	Description	Default
name	AgentName	Name for the new agent	required
result_type	type[Any] | str | ResponseDefinition | None	Optional type for structured responses: - None: Regular unstructured agent - type: Python type for validation - str: Name of response definition - ResponseDefinition: Complete response definition	None
**kwargs	Unpack[AgentKwargs]	Additional agent configuration	{}

Returns:

Type	Description
Agent[Any] | StructuredAgent[Any, Any]	Either a regular Agent or StructuredAgent depending on result_type

Source code in src/llmling_agent/delegation/pool.py

async def add_agent(
    self,
    name: AgentName,
    *,
    result_type: type[Any] | str | ResponseDefinition | None = None,
    **kwargs: Unpack[AgentKwargs],
) -> Agent[Any] | StructuredAgent[Any, Any]:
    """Add a new permanent agent to the pool.

    Args:
        name: Name for the new agent
        result_type: Optional type for structured responses:
            - None: Regular unstructured agent
            - type: Python type for validation
            - str: Name of response definition
            - ResponseDefinition: Complete response definition
        **kwargs: Additional agent configuration

    Returns:
        Either a regular Agent or StructuredAgent depending on result_type
    """
    from llmling_agent.agent import Agent

    agent: AnyAgent[Any, Any] = Agent(name=name, **kwargs)
    agent.tools.add_provider(self.mcp)
    agent = await self.exit_stack.enter_async_context(agent)
    # Convert to structured if needed
    if result_type is not None:
        agent = agent.to_structured(result_type)
    self.register(name, agent)
    return agent

cleanup async

cleanup()

Clean up all agents.

Source code in src/llmling_agent/delegation/pool.py

async def cleanup(self):
    """Clean up all agents."""
    await self.exit_stack.aclose()
    self.clear()

clone_agent async

clone_agent(
    agent: AgentName | Agent[TDeps],
    new_name: AgentName | None = None,
    *,
    system_prompts: list[str] | None = None,
    template_context: dict[str, Any] | None = None,
) -> Agent[TDeps]

clone_agent(
    agent: StructuredAgent[TDeps, TResult],
    new_name: AgentName | None = None,
    *,
    system_prompts: list[str] | None = None,
    template_context: dict[str, Any] | None = None,
) -> StructuredAgent[TDeps, TResult]

clone_agent(
    agent: AgentName | AnyAgent[TDeps, TAgentResult],
    new_name: AgentName | None = None,
    *,
    system_prompts: list[str] | None = None,
    template_context: dict[str, Any] | None = None,
) -> AnyAgent[TDeps, TAgentResult]

Create a copy of an agent.

Parameters:

Name	Type	Description	Default
agent	AgentName | AnyAgent[TDeps, TAgentResult]	Agent instance or name to clone	required
new_name	AgentName | None	Optional name for the clone	None
system_prompts	list[str] | None	Optional different prompts	None
template_context	dict[str, Any] | None	Variables for template rendering	None

Returns:

Type	Description
AnyAgent[TDeps, TAgentResult]	The new agent instance

Source code in src/llmling_agent/delegation/pool.py

async def clone_agent[TDeps, TAgentResult](
    self,
    agent: AgentName | AnyAgent[TDeps, TAgentResult],
    new_name: AgentName | None = None,
    *,
    system_prompts: list[str] | None = None,
    template_context: dict[str, Any] | None = None,
) -> AnyAgent[TDeps, TAgentResult]:
    """Create a copy of an agent.

    Args:
        agent: Agent instance or name to clone
        new_name: Optional name for the clone
        system_prompts: Optional different prompts
        template_context: Variables for template rendering

    Returns:
        The new agent instance
    """
    from llmling_agent.agent import Agent, StructuredAgent

    # Get original config
    if isinstance(agent, str):
        if agent not in self.manifest.agents:
            msg = f"Agent {agent} not found"
            raise KeyError(msg)
        config = self.manifest.agents[agent]
        original_agent: AnyAgent[Any, Any] = self.get_agent(agent)
    else:
        config = agent.context.config  # type: ignore
        original_agent = agent

    # Create new config
    new_config = config.model_copy(deep=True)

    # Apply overrides
    if system_prompts:
        new_config.system_prompts = system_prompts

    # Handle template rendering
    if template_context:
        new_config.system_prompts = new_config.render_system_prompts(template_context)

    # Create new agent with same runtime
    new_agent = Agent[TDeps](
        runtime=original_agent.runtime,
        context=original_agent.context,
        # result_type=original_agent.actual_type,
        provider=new_config.get_provider(),
        system_prompt=new_config.system_prompts,
        name=new_name or f"{config.name}_copy_{len(self.agents)}",
    )
    if isinstance(original_agent, StructuredAgent):
        new_agent = new_agent.to_structured(original_agent.actual_type)

    # Register in pool
    agent_name = new_agent.name
    self.manifest.agents[agent_name] = new_config
    self.register(agent_name, new_agent)
    return await self.exit_stack.enter_async_context(new_agent)

create_agent async

create_agent(
    name: AgentName,
    *,
    session: SessionIdType | SessionQuery = None,
    name_override: str | None = None,
) -> Agent[TPoolDeps]

create_agent(
    name: AgentName,
    *,
    deps: TCustomDeps,
    session: SessionIdType | SessionQuery = None,
    name_override: str | None = None,
) -> Agent[TCustomDeps]

create_agent(
    name: AgentName,
    *,
    return_type: type[TResult],
    session: SessionIdType | SessionQuery = None,
    name_override: str | None = None,
) -> StructuredAgent[TPoolDeps, TResult]

create_agent(
    name: AgentName,
    *,
    deps: TCustomDeps,
    return_type: type[TResult],
    session: SessionIdType | SessionQuery = None,
    name_override: str | None = None,
) -> StructuredAgent[TCustomDeps, TResult]

create_agent(
    name: AgentName,
    *,
    deps: Any | None = None,
    return_type: Any | None = None,
    session: SessionIdType | SessionQuery = None,
    name_override: str | None = None,
) -> AnyAgent[Any, Any]

Create a new agent instance from configuration.

Parameters:

Name	Type	Description	Default
name	AgentName	Name of the agent configuration to use	required
deps	Any | None	Optional custom dependencies (overrides pool deps)	None
return_type	Any | None	Optional type for structured responses	None
session	SessionIdType | SessionQuery	Optional session ID or query to recover conversation	None
name_override	str | None	Optional different name for this instance	None

Returns:

Type	Description
AnyAgent[Any, Any]	New agent instance with the specified configuration

Raises:

Type	Description
KeyError	If agent configuration not found
ValueError	If configuration is invalid

Source code in src/llmling_agent/delegation/pool.py

async def create_agent(
    self,
    name: AgentName,
    *,
    deps: Any | None = None,
    return_type: Any | None = None,
    session: SessionIdType | SessionQuery = None,
    name_override: str | None = None,
) -> AnyAgent[Any, Any]:
    """Create a new agent instance from configuration.

    Args:
        name: Name of the agent configuration to use
        deps: Optional custom dependencies (overrides pool deps)
        return_type: Optional type for structured responses
        session: Optional session ID or query to recover conversation
        name_override: Optional different name for this instance

    Returns:
        New agent instance with the specified configuration

    Raises:
        KeyError: If agent configuration not found
        ValueError: If configuration is invalid
    """
    if name not in self.manifest.agents:
        msg = f"Agent configuration {name!r} not found"
        raise KeyError(msg)

    # Use Manifest.get_agent for proper initialization
    final_deps = deps if deps is not None else self.shared_deps
    agent = self.manifest.get_agent(name, deps=final_deps)
    # Override name if requested
    if name_override:
        agent.name = name_override

    # Set pool reference
    agent.context.pool = self

    # Handle session if provided
    if session:
        agent.conversation.load_history_from_database(session=session)

    # Initialize agent through exit stack
    agent = await self.exit_stack.enter_async_context(agent)

    # Override structured configuration if provided
    if return_type is not None:
        return agent.to_structured(return_type)

    return agent

create_team

create_team(agents: Sequence[str]) -> Team[TPoolDeps]

create_team(
    agents: Sequence[MessageNode[TDeps, Any]],
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> Team[TDeps]

create_team(
    agents: Sequence[AgentName | MessageNode[Any, Any]],
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> Team[Any]

create_team(
    agents: Sequence[AgentName | MessageNode[Any, Any]] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> Team[Any]

Create a group from agent names or instances.

Parameters:

Name	Type	Description	Default
agents	Sequence[AgentName | MessageNode[Any, Any]] | None	List of agent names or instances (all if None)	None
name	str | None	Optional name for the team	None
description	str | None	Optional description for the team	None
shared_prompt	str | None	Optional prompt for all agents	None
picker	AnyAgent[Any, Any] | None	Agent to use for picking agents	None
num_picks	int | None	Number of agents to pick	None
pick_prompt	str | None	Prompt to use for picking agents	None

Source code in src/llmling_agent/delegation/pool.py

def create_team(
    self,
    agents: Sequence[AgentName | MessageNode[Any, Any]] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> Team[Any]:
    """Create a group from agent names or instances.

    Args:
        agents: List of agent names or instances (all if None)
        name: Optional name for the team
        description: Optional description for the team
        shared_prompt: Optional prompt for all agents
        picker: Agent to use for picking agents
        num_picks: Number of agents to pick
        pick_prompt: Prompt to use for picking agents
    """
    from llmling_agent.delegation.team import Team

    if agents is None:
        agents = list(self.agents.keys())

    # First resolve/configure agents
    resolved_agents: list[MessageNode[Any, Any]] = []
    for agent in agents:
        if isinstance(agent, str):
            agent = self.get_agent(agent)
        resolved_agents.append(agent)

    team = Team(
        name=name,
        description=description,
        agents=resolved_agents,
        shared_prompt=shared_prompt,
        picker=picker,
        num_picks=num_picks,
        pick_prompt=pick_prompt,
    )
    if name:
        self[name] = team
    return team

create_team_run

create_team_run(
    agents: Sequence[str],
    validator: MessageNode[Any, TResult] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> TeamRun[TPoolDeps, TResult]

create_team_run(
    agents: Sequence[MessageNode[TDeps, Any]],
    validator: MessageNode[Any, TResult] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> TeamRun[TDeps, TResult]

create_team_run(
    agents: Sequence[AgentName | MessageNode[Any, Any]],
    validator: MessageNode[Any, TResult] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> TeamRun[Any, TResult]

create_team_run(
    agents: Sequence[AgentName | MessageNode[Any, Any]] | None = None,
    validator: MessageNode[Any, TResult] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> TeamRun[Any, TResult]

Create a a sequential TeamRun from a list of Agents.

Parameters:

Name	Type	Description	Default
agents	Sequence[AgentName | MessageNode[Any, Any]] | None	List of agent names or team/agent instances (all if None)	None
validator	MessageNode[Any, TResult] | None	Node to validate the results of the TeamRun	None
name	str | None	Optional name for the team	None
description	str | None	Optional description for the team	None
shared_prompt	str | None	Optional prompt for all agents	None
picker	AnyAgent[Any, Any] | None	Agent to use for picking agents	None
num_picks	int | None	Number of agents to pick	None
pick_prompt	str | None	Prompt to use for picking agents	None

Source code in src/llmling_agent/delegation/pool.py

def create_team_run(
    self,
    agents: Sequence[AgentName | MessageNode[Any, Any]] | None = None,
    validator: MessageNode[Any, TResult] | None = None,
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
) -> TeamRun[Any, TResult]:
    """Create a a sequential TeamRun from a list of Agents.

    Args:
        agents: List of agent names or team/agent instances (all if None)
        validator: Node to validate the results of the TeamRun
        name: Optional name for the team
        description: Optional description for the team
        shared_prompt: Optional prompt for all agents
        picker: Agent to use for picking agents
        num_picks: Number of agents to pick
        pick_prompt: Prompt to use for picking agents
    """
    from llmling_agent.delegation.teamrun import TeamRun

    if agents is None:
        agents = list(self.agents.keys())

    # First resolve/configure agents
    resolved_agents: list[MessageNode[Any, Any]] = []
    for agent in agents:
        if isinstance(agent, str):
            agent = self.get_agent(agent)
        resolved_agents.append(agent)
    team = TeamRun(
        resolved_agents,
        name=name,
        description=description,
        validator=validator,
        shared_prompt=shared_prompt,
        picker=picker,
        num_picks=num_picks,
        pick_prompt=pick_prompt,
    )
    if name:
        self[name] = team
    return team

get_agent

get_agent(
    agent: AgentName | Agent[Any],
    *,
    model_override: str | None = None,
    session: SessionIdType | SessionQuery = None,
) -> Agent[TPoolDeps]

get_agent(
    agent: AgentName | Agent[Any],
    *,
    return_type: type[TResult],
    model_override: str | None = None,
    session: SessionIdType | SessionQuery = None,
) -> StructuredAgent[TPoolDeps, TResult]

get_agent(
    agent: AgentName | Agent[Any],
    *,
    deps: TCustomDeps,
    model_override: str | None = None,
    session: SessionIdType | SessionQuery = None,
) -> Agent[TCustomDeps]

get_agent(
    agent: AgentName | Agent[Any],
    *,
    deps: TCustomDeps,
    return_type: type[TResult],
    model_override: str | None = None,
    session: SessionIdType | SessionQuery = None,
) -> StructuredAgent[TCustomDeps, TResult]

get_agent(
    agent: AgentName | Agent[Any],
    *,
    deps: Any | None = None,
    return_type: Any | None = None,
    model_override: str | None = None,
    session: SessionIdType | SessionQuery = None,
) -> AnyAgent[Any, Any]

Get or configure an agent from the pool.

This method provides flexible agent configuration with dependency injection: - Without deps: Agent uses pool's shared dependencies - With deps: Agent uses provided custom dependencies - With return_type: Returns a StructuredAgent with type validation

Parameters:

Name	Type	Description	Default
agent	AgentName | Agent[Any]	Either agent name or instance	required
deps	Any | None	Optional custom dependencies (overrides shared deps)	None
return_type	Any | None	Optional type for structured responses	None
model_override	str | None	Optional model override	None
session	SessionIdType | SessionQuery	Optional session ID or query to recover conversation	None

Returns:

Name	Type	Description
Either	AnyAgent[Any, Any]
	AnyAgent[Any, Any]	Agent[TPoolDeps] when using pool's shared deps
	AnyAgent[Any, Any]	Agent[TCustomDeps] when custom deps provided
	AnyAgent[Any, Any]	StructuredAgent when return_type provided

Raises:

Type	Description
KeyError	If agent name not found
ValueError	If configuration is invalid

Source code in src/llmling_agent/delegation/pool.py

def get_agent(
    self,
    agent: AgentName | Agent[Any],
    *,
    deps: Any | None = None,
    return_type: Any | None = None,
    model_override: str | None = None,
    session: SessionIdType | SessionQuery = None,
) -> AnyAgent[Any, Any]:
    """Get or configure an agent from the pool.

    This method provides flexible agent configuration with dependency injection:
    - Without deps: Agent uses pool's shared dependencies
    - With deps: Agent uses provided custom dependencies
    - With return_type: Returns a StructuredAgent with type validation

    Args:
        agent: Either agent name or instance
        deps: Optional custom dependencies (overrides shared deps)
        return_type: Optional type for structured responses
        model_override: Optional model override
        session: Optional session ID or query to recover conversation

    Returns:
        Either:
        - Agent[TPoolDeps] when using pool's shared deps
        - Agent[TCustomDeps] when custom deps provided
        - StructuredAgent when return_type provided

    Raises:
        KeyError: If agent name not found
        ValueError: If configuration is invalid
    """
    from llmling_agent.agent import Agent
    from llmling_agent.agent.context import AgentContext

    # Get base agent
    base = agent if isinstance(agent, Agent) else self.agents[agent]

    # Setup context and dependencies
    if base.context is None:
        base.context = AgentContext[Any].create_default(base.name)

    # Use custom deps if provided, otherwise use shared deps
    base.context.data = deps if deps is not None else self.shared_deps
    base.context.pool = self

    # Apply overrides
    if model_override:
        base.set_model(model_override)

    if session:
        base.conversation.load_history_from_database(session=session)

    # Convert to structured if needed
    if return_type is not None:
        return base.to_structured(return_type)

    return base

get_mermaid_diagram

get_mermaid_diagram(include_details: bool = True) -> str

Generate mermaid flowchart of all agents and their connections.

Parameters:

Name	Type	Description	Default
include_details	bool	Whether to show connection details (types, queues, etc)	True

Source code in src/llmling_agent/delegation/pool.py

def get_mermaid_diagram(
    self,
    include_details: bool = True,
) -> str:
    """Generate mermaid flowchart of all agents and their connections.

    Args:
        include_details: Whether to show connection details (types, queues, etc)
    """
    lines = ["flowchart LR"]

    # Add all agents as nodes
    for name in self.agents:
        lines.append(f"    {name}[{name}]")  # noqa: PERF401

    # Add all connections as edges
    for agent in self.agents.values():
        connections = agent.connections.get_connections()
        for talk in connections:
            talk = cast(Talk[Any], talk)  # help mypy understand it's a Talk
            source = talk.source.name
            for target in talk.targets:
                if include_details:
                    details: list[str] = []
                    details.append(talk.connection_type)
                    if talk.queued:
                        details.append(f"queued({talk.queue_strategy})")
                    if fn := talk.filter_condition:  # type: ignore
                        details.append(f"filter:{fn.__name__}")
                    if fn := talk.stop_condition:  # type: ignore
                        details.append(f"stop:{fn.__name__}")
                    if fn := talk.exit_condition:  # type: ignore
                        details.append(f"exit:{fn.__name__}")

                    label = f"|{' '.join(details)}|" if details else ""
                    lines.append(f"    {source}--{label}-->{target.name}")
                else:
                    lines.append(f"    {source}-->{target.name}")

    return "\n".join(lines)

list_nodes

list_nodes() -> list[str]

List available agent names.

Source code in src/llmling_agent/delegation/pool.py

def list_nodes(self) -> list[str]:
    """List available agent names."""
    return list(self.list_items())

run_event_loop async

run_event_loop()

Run pool in event-watching mode until interrupted.

Source code in src/llmling_agent/delegation/pool.py

async def run_event_loop(self):
    """Run pool in event-watching mode until interrupted."""
    import sys

    print("Starting event watch mode...")
    print("Active nodes: ", ", ".join(self.list_nodes()))
    print("Press Ctrl+C to stop")

    stop_event = asyncio.Event()

    if sys.platform != "win32":
        # Unix: Use signal handlers
        loop = asyncio.get_running_loop()
        for sig in (signal.SIGINT, signal.SIGTERM):
            loop.add_signal_handler(sig, stop_event.set)
        while True:
            await asyncio.sleep(1)
    else:
        # Windows: Use keyboard interrupt
        with suppress(KeyboardInterrupt):
            while True:
                await asyncio.sleep(1)

setup_agent_workers

setup_agent_workers(agent: AnyAgent[Any, Any])

Set up workers for an agent from configuration.

Source code in src/llmling_agent/delegation/pool.py

def setup_agent_workers(self, agent: AnyAgent[Any, Any]):
    """Set up workers for an agent from configuration."""
    for worker_config in agent.context.config.workers:
        try:
            worker = self.nodes[worker_config.name]
            match worker_config:
                case TeamWorkerConfig():
                    agent.register_worker(worker)
                case AgentWorkerConfig():
                    agent.register_worker(
                        worker,
                        name=worker_config.name,
                        reset_history_on_run=worker_config.reset_history_on_run,
                        pass_message_history=worker_config.pass_message_history,
                        share_context=worker_config.share_context,
                    )
        except KeyError as e:
            msg = f"Worker agent {worker_config.name!r} not found"
            raise ValueError(msg) from e

track_message_flow async

track_message_flow() -> AsyncIterator[MessageFlowTracker]

Track message flow during a context.

Source code in src/llmling_agent/delegation/pool.py

@asynccontextmanager
async def track_message_flow(self) -> AsyncIterator[MessageFlowTracker]:
    """Track message flow during a context."""
    tracker = MessageFlowTracker()
    self.connection_registry.message_flow.connect(tracker.track)
    try:
        yield tracker
    finally:
        self.connection_registry.message_flow.disconnect(tracker.track)

AgentsManifest

Bases: ConfigModel

Complete agent configuration manifest defining all available agents.

This is the root configuration that: - Defines available response types (both inline and imported) - Configures all agent instances and their settings - Sets up custom role definitions and capabilities - Manages environment configurations

A single manifest can define multiple agents that can work independently or collaborate through the orchestrator.

Source code in src/llmling_agent/models/manifest.py

class AgentsManifest(ConfigModel):
    """Complete agent configuration manifest defining all available agents.

    This is the root configuration that:
    - Defines available response types (both inline and imported)
    - Configures all agent instances and their settings
    - Sets up custom role definitions and capabilities
    - Manages environment configurations

    A single manifest can define multiple agents that can work independently
    or collaborate through the orchestrator.
    """

    INHERIT: str | list[str] | None = None
    """Inheritance references."""

    resources: dict[str, ResourceConfig | str] = Field(default_factory=dict)
    """Resource configurations defining available filesystems.

    Supports both full config and URI shorthand:
        resources:
          docs: "file://./docs"  # shorthand
          data:  # full config
            type: "source"
            uri: "s3://bucket/data"
            cached: true
    """

    ui: UIConfig = Field(default_factory=StdlibUIConfig)
    """UI configuration."""

    agents: dict[str, AgentConfig] = Field(default_factory=dict)
    """Mapping of agent IDs to their configurations"""

    teams: dict[str, TeamConfig] = Field(default_factory=dict)
    """Mapping of team IDs to their configurations"""

    storage: StorageConfig = Field(default_factory=StorageConfig)
    """Storage provider configuration."""

    observability: ObservabilityConfig = Field(default_factory=ObservabilityConfig)
    """Observability provider configuration."""

    conversion: ConversionConfig = Field(default_factory=ConversionConfig)
    """Document conversion configuration."""

    responses: dict[str, ResponseDefinition] = Field(default_factory=dict)
    """Mapping of response names to their definitions"""

    jobs: dict[str, Job] = Field(default_factory=dict)
    """Pre-defined jobs, ready to be used by nodes."""

    mcp_servers: list[str | MCPServerConfig] = Field(default_factory=list)
    """List of MCP server configurations:

    These MCP servers are used to provide tools and other resources to the nodes.
    """
    pool_server: PoolServerConfig = Field(default_factory=PoolServerConfig)
    """Pool server configuration.

    This MCP server configuration is used for the pool MCP server,
    which exposes pool functionality to other applications / clients."""

    prompts: PromptConfig = Field(default_factory=PromptConfig)

    model_config = ConfigDict(use_attribute_docstrings=True, extra="forbid")

    @model_validator(mode="before")
    @classmethod
    def normalize_workers(cls, data: dict[str, Any]) -> dict[str, Any]:
        """Convert string workers to appropriate WorkerConfig for all agents."""
        teams = data.get("teams", {})
        agents = data.get("agents", {})

        # Process workers for all agents that have them
        for agent_name, agent_config in agents.items():
            if isinstance(agent_config, dict):
                workers = agent_config.get("workers", [])
            else:
                workers = agent_config.workers

            if workers:
                normalized: list[BaseWorkerConfig] = []

                for worker in workers:
                    match worker:
                        case str() as name:
                            # Determine type based on presence in teams/agents
                            if name in teams:
                                normalized.append(TeamWorkerConfig(name=name))
                            elif name in agents:
                                normalized.append(AgentWorkerConfig(name=name))
                            else:
                                # Default to agent if type can't be determined
                                normalized.append(AgentWorkerConfig(name=name))

                        case dict() as config:
                            # If type is explicitly specified, use it
                            if worker_type := config.get("type"):
                                match worker_type:
                                    case "team":
                                        normalized.append(TeamWorkerConfig(**config))
                                    case "agent":
                                        normalized.append(AgentWorkerConfig(**config))
                                    case _:
                                        msg = f"Invalid worker type: {worker_type}"
                                        raise ValueError(msg)
                            else:
                                # Determine type based on worker name
                                worker_name = config.get("name")
                                if not worker_name:
                                    msg = "Worker config missing name"
                                    raise ValueError(msg)

                                if worker_name in teams:
                                    normalized.append(TeamWorkerConfig(**config))
                                else:
                                    normalized.append(AgentWorkerConfig(**config))

                        case BaseWorkerConfig():  # Already normalized
                            normalized.append(worker)

                        case _:
                            msg = f"Invalid worker configuration: {worker}"
                            raise ValueError(msg)

                if isinstance(agent_config, dict):
                    agent_config["workers"] = normalized
                else:
                    # Need to create a new dict with updated workers
                    agent_dict = agent_config.model_dump()
                    agent_dict["workers"] = normalized
                    agents[agent_name] = agent_dict

        return data

    @cached_property
    def resource_registry(self) -> ResourceRegistry:
        """Get registry with all configured resources."""
        registry = ResourceRegistry()
        for name, config in self.resources.items():
            if isinstance(config, str):
                # Convert URI shorthand to SourceResourceConfig
                config = SourceResourceConfig(uri=config)
            registry.register_from_config(name, config)
        return registry

    def clone_agent_config(
        self,
        name: str,
        new_name: str | None = None,
        *,
        template_context: dict[str, Any] | None = None,
        **overrides: Any,
    ) -> str:
        """Create a copy of an agent configuration.

        Args:
            name: Name of agent to clone
            new_name: Optional new name (auto-generated if None)
            template_context: Variables for template rendering
            **overrides: Configuration overrides for the clone

        Returns:
            Name of the new agent

        Raises:
            KeyError: If original agent not found
            ValueError: If new name already exists or if overrides invalid
        """
        if name not in self.agents:
            msg = f"Agent {name} not found"
            raise KeyError(msg)

        actual_name = new_name or f"{name}_copy_{len(self.agents)}"
        if actual_name in self.agents:
            msg = f"Agent {actual_name} already exists"
            raise ValueError(msg)

        # Deep copy the configuration
        config = self.agents[name].model_copy(deep=True)

        # Apply overrides
        for key, value in overrides.items():
            if not hasattr(config, key):
                msg = f"Invalid override: {key}"
                raise ValueError(msg)
            setattr(config, key, value)

        # Handle template rendering if context provided
        if template_context:
            # Apply name from context if not explicitly overridden
            if "name" in template_context and "name" not in overrides:
                config.name = template_context["name"]

            # Render system prompts
            config.system_prompts = config.render_system_prompts(template_context)

        self.agents[actual_name] = config
        return actual_name

    @model_validator(mode="before")
    @classmethod
    def resolve_inheritance(cls, data: dict) -> dict:
        """Resolve agent inheritance chains."""
        nodes = data.get("agents", {})
        resolved: dict[str, dict] = {}
        seen: set[str] = set()

        def resolve_node(name: str) -> dict:
            if name in resolved:
                return resolved[name]

            if name in seen:
                msg = f"Circular inheritance detected: {name}"
                raise ValueError(msg)

            seen.add(name)
            config = (
                nodes[name].model_copy()
                if hasattr(nodes[name], "model_copy")
                else nodes[name].copy()
            )
            inherit = (
                config.get("inherits") if isinstance(config, dict) else config.inherits
            )
            if inherit:
                if inherit not in nodes:
                    msg = f"Parent agent {inherit} not found"
                    raise ValueError(msg)

                # Get resolved parent config
                parent = resolve_node(inherit)
                # Merge parent with child (child overrides parent)
                merged = parent.copy()
                merged.update(config)
                config = merged

            seen.remove(name)
            resolved[name] = config
            return config

        # Resolve all nodes
        for name in nodes:
            resolved[name] = resolve_node(name)

        # Update nodes with resolved configs
        data["agents"] = resolved
        return data

    @model_validator(mode="after")
    def set_instrument_libraries(self) -> Self:
        """Auto-set libraries to instrument based on used providers."""
        if (
            not self.observability.enabled
            or self.observability.instrument_libraries is not None
        ):
            return self
        self.observability.instrument_libraries = list(self.get_used_providers())
        return self

    @property
    def node_names(self) -> list[str]:
        """Get list of all agent and team names."""
        return list(self.agents.keys()) + list(self.teams.keys())

    @property
    def nodes(self) -> dict[str, Any]:
        """Get all agent and team configurations."""
        return {**self.agents, **self.teams}

    def get_mcp_servers(self) -> list[MCPServerConfig]:
        """Get processed MCP server configurations.

        Converts string entries to StdioMCPServerConfig configs by splitting
        into command and arguments.

        Returns:
            List of MCPServerConfig instances

        Raises:
            ValueError: If string entry is empty
        """
        configs: list[MCPServerConfig] = []

        for server in self.mcp_servers:
            match server:
                case str():
                    parts = server.split()
                    if not parts:
                        msg = "Empty MCP server command"
                        raise ValueError(msg)

                    configs.append(StdioMCPServerConfig(command=parts[0], args=parts[1:]))
                case BaseMCPServerConfig():
                    configs.append(server)

        return configs

    @cached_property
    def prompt_manager(self) -> PromptManager:
        """Get prompt manager for this manifest."""
        from llmling_agent.prompts.manager import PromptManager

        return PromptManager(self.prompts)

    # @model_validator(mode="after")
    # def validate_response_types(self) -> AgentsManifest:
    #     """Ensure all agent result_types exist in responses or are inline."""
    #     for agent_id, agent in self.agents.items():
    #         if (
    #             isinstance(agent.result_type, str)
    #             and agent.result_type not in self.responses
    #         ):
    #             msg = f"'{agent.result_type=}' for '{agent_id=}' not found in responses"
    #             raise ValueError(msg)
    #     return self

    def get_agent[TAgentDeps](
        self, name: str, deps: TAgentDeps | None = None
    ) -> AnyAgent[TAgentDeps, Any]:
        from llmling import RuntimeConfig

        from llmling_agent import Agent, AgentContext

        config = self.agents[name]
        # Create runtime without async context
        cfg = config.get_config()
        runtime = RuntimeConfig.from_config(cfg)

        # Create context with config path and capabilities
        context = AgentContext[TAgentDeps](
            node_name=name,
            data=deps,
            capabilities=config.capabilities,
            definition=self,
            config=config,
            runtime=runtime,
            # pool=self,
            # confirmation_callback=confirmation_callback,
        )

        sys_prompts = config.system_prompts.copy()
        # Library prompts
        if config.library_system_prompts:
            for prompt_ref in config.library_system_prompts:
                try:
                    content = self.prompt_manager.get_sync(prompt_ref)
                    sys_prompts.append(content)
                except Exception as e:
                    msg = f"Failed to load library prompt {prompt_ref!r} for agent {name}"
                    logger.exception(msg)
                    raise ValueError(msg) from e
        # Create agent with runtime and context
        agent = Agent[Any](
            runtime=runtime,
            context=context,
            provider=config.get_provider(),
            system_prompt=sys_prompts,
            name=name,
            description=config.description,
            retries=config.retries,
            session=config.get_session_config(),
            result_retries=config.result_retries,
            end_strategy=config.end_strategy,
            capabilities=config.capabilities,
            debug=config.debug,
            # name=config.name or name,
        )
        if result_type := self.get_result_type(name):
            return agent.to_structured(result_type)
        return agent

    def get_used_providers(self) -> set[str]:
        """Get all providers configured in this manifest."""
        providers = set[str]()

        for agent_config in self.agents.values():
            match agent_config.provider:
                case "pydantic_ai":
                    providers.add("pydantic_ai")
                case "litellm":
                    providers.add("litellm")
                case BaseProviderConfig():
                    providers.add(agent_config.provider.type)
        return providers

    @classmethod
    def from_file(cls, path: StrPath) -> Self:
        """Load agent configuration from YAML file.

        Args:
            path: Path to the configuration file

        Returns:
            Loaded agent definition

        Raises:
            ValueError: If loading fails
        """
        import yamling

        try:
            data = yamling.load_yaml_file(path, resolve_inherit=True)
            agent_def = cls.model_validate(data)
            # Update all agents with the config file path and ensure names
            agents = {
                name: config.model_copy(update={"config_file_path": str(path)})
                for name, config in agent_def.agents.items()
            }
            return agent_def.model_copy(update={"agents": agents})
        except Exception as exc:
            msg = f"Failed to load agent config from {path}"
            raise ValueError(msg) from exc

    @cached_property
    def pool(self) -> AgentPool:
        """Create an agent pool from this manifest.

        Returns:
            Configured agent pool
        """
        from llmling_agent import AgentPool

        return AgentPool(manifest=self)

    def get_result_type(self, agent_name: str) -> type[Any] | None:
        """Get the resolved result type for an agent.

        Returns None if no result type is configured.
        """
        agent_config = self.agents[agent_name]
        if not agent_config.result_type:
            return None
        logger.debug("Building response model for %r", agent_config.result_type)
        if isinstance(agent_config.result_type, str):
            response_def = self.responses[agent_config.result_type]
            return response_def.create_model()  # type: ignore
        return agent_config.result_type.create_model()  # type: ignore

INHERIT class-attribute instance-attribute

INHERIT: str | list[str] | None = None

Inheritance references.

agents class-attribute instance-attribute

agents: dict[str, AgentConfig] = Field(default_factory=dict)

Mapping of agent IDs to their configurations

conversion class-attribute instance-attribute

conversion: ConversionConfig = Field(default_factory=ConversionConfig)

Document conversion configuration.

jobs class-attribute instance-attribute

jobs: dict[str, Job] = Field(default_factory=dict)

Pre-defined jobs, ready to be used by nodes.

mcp_servers class-attribute instance-attribute

mcp_servers: list[str | MCPServerConfig] = Field(default_factory=list)

List of MCP server configurations:

These MCP servers are used to provide tools and other resources to the nodes.

node_names property

node_names: list[str]

Get list of all agent and team names.

nodes property

nodes: dict[str, Any]

Get all agent and team configurations.

observability class-attribute instance-attribute

observability: ObservabilityConfig = Field(default_factory=ObservabilityConfig)

Observability provider configuration.

pool cached property

pool: AgentPool

Create an agent pool from this manifest.

Returns:

Type	Description
AgentPool	Configured agent pool

pool_server class-attribute instance-attribute

pool_server: PoolServerConfig = Field(default_factory=PoolServerConfig)

Pool server configuration.

This MCP server configuration is used for the pool MCP server, which exposes pool functionality to other applications / clients.

prompt_manager cached property

prompt_manager: PromptManager

Get prompt manager for this manifest.

resource_registry cached property

resource_registry: ResourceRegistry

Get registry with all configured resources.

resources class-attribute instance-attribute

resources: dict[str, ResourceConfig | str] = Field(default_factory=dict)

Resource configurations defining available filesystems.

Supports both full config and URI shorthand

resources: docs: "file://./docs" # shorthand data: # full config type: "source" uri: "s3://bucket/data" cached: true

responses class-attribute instance-attribute

responses: dict[str, ResponseDefinition] = Field(default_factory=dict)

Mapping of response names to their definitions

storage class-attribute instance-attribute

storage: StorageConfig = Field(default_factory=StorageConfig)

Storage provider configuration.

teams class-attribute instance-attribute

teams: dict[str, TeamConfig] = Field(default_factory=dict)

Mapping of team IDs to their configurations

ui class-attribute instance-attribute

ui: UIConfig = Field(default_factory=StdlibUIConfig)

UI configuration.

clone_agent_config

clone_agent_config(
    name: str,
    new_name: str | None = None,
    *,
    template_context: dict[str, Any] | None = None,
    **overrides: Any,
) -> str

Create a copy of an agent configuration.

Parameters:

Name	Type	Description	Default
name	str	Name of agent to clone	required
new_name	str | None	Optional new name (auto-generated if None)	None
template_context	dict[str, Any] | None	Variables for template rendering	None
**overrides	Any	Configuration overrides for the clone	{}

Returns:

Type	Description
str	Name of the new agent

Raises:

Type	Description
KeyError	If original agent not found
ValueError	If new name already exists or if overrides invalid

Source code in src/llmling_agent/models/manifest.py

def clone_agent_config(
    self,
    name: str,
    new_name: str | None = None,
    *,
    template_context: dict[str, Any] | None = None,
    **overrides: Any,
) -> str:
    """Create a copy of an agent configuration.

    Args:
        name: Name of agent to clone
        new_name: Optional new name (auto-generated if None)
        template_context: Variables for template rendering
        **overrides: Configuration overrides for the clone

    Returns:
        Name of the new agent

    Raises:
        KeyError: If original agent not found
        ValueError: If new name already exists or if overrides invalid
    """
    if name not in self.agents:
        msg = f"Agent {name} not found"
        raise KeyError(msg)

    actual_name = new_name or f"{name}_copy_{len(self.agents)}"
    if actual_name in self.agents:
        msg = f"Agent {actual_name} already exists"
        raise ValueError(msg)

    # Deep copy the configuration
    config = self.agents[name].model_copy(deep=True)

    # Apply overrides
    for key, value in overrides.items():
        if not hasattr(config, key):
            msg = f"Invalid override: {key}"
            raise ValueError(msg)
        setattr(config, key, value)

    # Handle template rendering if context provided
    if template_context:
        # Apply name from context if not explicitly overridden
        if "name" in template_context and "name" not in overrides:
            config.name = template_context["name"]

        # Render system prompts
        config.system_prompts = config.render_system_prompts(template_context)

    self.agents[actual_name] = config
    return actual_name

from_file classmethod

from_file(path: StrPath) -> Self

Load agent configuration from YAML file.

Parameters:

Name	Type	Description	Default
path	StrPath	Path to the configuration file	required

Returns:

Type	Description
Self	Loaded agent definition

Raises:

Type	Description
ValueError	If loading fails

Source code in src/llmling_agent/models/manifest.py

@classmethod
def from_file(cls, path: StrPath) -> Self:
    """Load agent configuration from YAML file.

    Args:
        path: Path to the configuration file

    Returns:
        Loaded agent definition

    Raises:
        ValueError: If loading fails
    """
    import yamling

    try:
        data = yamling.load_yaml_file(path, resolve_inherit=True)
        agent_def = cls.model_validate(data)
        # Update all agents with the config file path and ensure names
        agents = {
            name: config.model_copy(update={"config_file_path": str(path)})
            for name, config in agent_def.agents.items()
        }
        return agent_def.model_copy(update={"agents": agents})
    except Exception as exc:
        msg = f"Failed to load agent config from {path}"
        raise ValueError(msg) from exc

get_mcp_servers

get_mcp_servers() -> list[MCPServerConfig]

Get processed MCP server configurations.

Converts string entries to StdioMCPServerConfig configs by splitting into command and arguments.

Returns:

Type	Description
list[MCPServerConfig]	List of MCPServerConfig instances

Raises:

Type	Description
ValueError	If string entry is empty

Source code in src/llmling_agent/models/manifest.py

def get_mcp_servers(self) -> list[MCPServerConfig]:
    """Get processed MCP server configurations.

    Converts string entries to StdioMCPServerConfig configs by splitting
    into command and arguments.

    Returns:
        List of MCPServerConfig instances

    Raises:
        ValueError: If string entry is empty
    """
    configs: list[MCPServerConfig] = []

    for server in self.mcp_servers:
        match server:
            case str():
                parts = server.split()
                if not parts:
                    msg = "Empty MCP server command"
                    raise ValueError(msg)

                configs.append(StdioMCPServerConfig(command=parts[0], args=parts[1:]))
            case BaseMCPServerConfig():
                configs.append(server)

    return configs

get_result_type

get_result_type(agent_name: str) -> type[Any] | None

Get the resolved result type for an agent.

Returns None if no result type is configured.

Source code in src/llmling_agent/models/manifest.py

def get_result_type(self, agent_name: str) -> type[Any] | None:
    """Get the resolved result type for an agent.

    Returns None if no result type is configured.
    """
    agent_config = self.agents[agent_name]
    if not agent_config.result_type:
        return None
    logger.debug("Building response model for %r", agent_config.result_type)
    if isinstance(agent_config.result_type, str):
        response_def = self.responses[agent_config.result_type]
        return response_def.create_model()  # type: ignore
    return agent_config.result_type.create_model()  # type: ignore

get_used_providers

get_used_providers() -> set[str]

Get all providers configured in this manifest.

Source code in src/llmling_agent/models/manifest.py

def get_used_providers(self) -> set[str]:
    """Get all providers configured in this manifest."""
    providers = set[str]()

    for agent_config in self.agents.values():
        match agent_config.provider:
            case "pydantic_ai":
                providers.add("pydantic_ai")
            case "litellm":
                providers.add("litellm")
            case BaseProviderConfig():
                providers.add(agent_config.provider.type)
    return providers

normalize_workers classmethod

normalize_workers(data: dict[str, Any]) -> dict[str, Any]

Convert string workers to appropriate WorkerConfig for all agents.

Source code in src/llmling_agent/models/manifest.py

@model_validator(mode="before")
@classmethod
def normalize_workers(cls, data: dict[str, Any]) -> dict[str, Any]:
    """Convert string workers to appropriate WorkerConfig for all agents."""
    teams = data.get("teams", {})
    agents = data.get("agents", {})

    # Process workers for all agents that have them
    for agent_name, agent_config in agents.items():
        if isinstance(agent_config, dict):
            workers = agent_config.get("workers", [])
        else:
            workers = agent_config.workers

        if workers:
            normalized: list[BaseWorkerConfig] = []

            for worker in workers:
                match worker:
                    case str() as name:
                        # Determine type based on presence in teams/agents
                        if name in teams:
                            normalized.append(TeamWorkerConfig(name=name))
                        elif name in agents:
                            normalized.append(AgentWorkerConfig(name=name))
                        else:
                            # Default to agent if type can't be determined
                            normalized.append(AgentWorkerConfig(name=name))

                    case dict() as config:
                        # If type is explicitly specified, use it
                        if worker_type := config.get("type"):
                            match worker_type:
                                case "team":
                                    normalized.append(TeamWorkerConfig(**config))
                                case "agent":
                                    normalized.append(AgentWorkerConfig(**config))
                                case _:
                                    msg = f"Invalid worker type: {worker_type}"
                                    raise ValueError(msg)
                        else:
                            # Determine type based on worker name
                            worker_name = config.get("name")
                            if not worker_name:
                                msg = "Worker config missing name"
                                raise ValueError(msg)

                            if worker_name in teams:
                                normalized.append(TeamWorkerConfig(**config))
                            else:
                                normalized.append(AgentWorkerConfig(**config))

                    case BaseWorkerConfig():  # Already normalized
                        normalized.append(worker)

                    case _:
                        msg = f"Invalid worker configuration: {worker}"
                        raise ValueError(msg)

            if isinstance(agent_config, dict):
                agent_config["workers"] = normalized
            else:
                # Need to create a new dict with updated workers
                agent_dict = agent_config.model_dump()
                agent_dict["workers"] = normalized
                agents[agent_name] = agent_dict

    return data

resolve_inheritance classmethod

resolve_inheritance(data: dict) -> dict

Resolve agent inheritance chains.

Source code in src/llmling_agent/models/manifest.py

@model_validator(mode="before")
@classmethod
def resolve_inheritance(cls, data: dict) -> dict:
    """Resolve agent inheritance chains."""
    nodes = data.get("agents", {})
    resolved: dict[str, dict] = {}
    seen: set[str] = set()

    def resolve_node(name: str) -> dict:
        if name in resolved:
            return resolved[name]

        if name in seen:
            msg = f"Circular inheritance detected: {name}"
            raise ValueError(msg)

        seen.add(name)
        config = (
            nodes[name].model_copy()
            if hasattr(nodes[name], "model_copy")
            else nodes[name].copy()
        )
        inherit = (
            config.get("inherits") if isinstance(config, dict) else config.inherits
        )
        if inherit:
            if inherit not in nodes:
                msg = f"Parent agent {inherit} not found"
                raise ValueError(msg)

            # Get resolved parent config
            parent = resolve_node(inherit)
            # Merge parent with child (child overrides parent)
            merged = parent.copy()
            merged.update(config)
            config = merged

        seen.remove(name)
        resolved[name] = config
        return config

    # Resolve all nodes
    for name in nodes:
        resolved[name] = resolve_node(name)

    # Update nodes with resolved configs
    data["agents"] = resolved
    return data

set_instrument_libraries

set_instrument_libraries() -> Self

Auto-set libraries to instrument based on used providers.

Source code in src/llmling_agent/models/manifest.py

@model_validator(mode="after")
def set_instrument_libraries(self) -> Self:
    """Auto-set libraries to instrument based on used providers."""
    if (
        not self.observability.enabled
        or self.observability.instrument_libraries is not None
    ):
        return self
    self.observability.instrument_libraries = list(self.get_used_providers())
    return self

AudioBase64Content

Bases: AudioContent

Audio from base64 data.

Source code in src/llmling_agent/models/content.py

class AudioBase64Content(AudioContent):
    """Audio from base64 data."""

    type: Literal["audio_base64"] = Field("audio_base64", init=False)
    """Base64-encoded audio."""

    data: str
    """Audio data in base64 format."""

    format: str | None = None  # mp3, wav, etc
    """Audio format."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for audio models."""
        data_url = f"data:audio/{self.format or 'mp3'};base64,{self.data}"
        content = {"url": data_url, "format": self.format or "auto"}
        return {"type": "audio", "audio": content}

    @classmethod
    def from_bytes(cls, data: bytes, audio_format: str = "mp3") -> Self:
        """Create from raw bytes."""
        return cls(data=base64.b64encode(data).decode(), format=audio_format)

    @classmethod
    def from_path(cls, path: StrPath) -> Self:
        """Create from file path with auto format detection."""
        import mimetypes

        from upath import UPath

        path_obj = UPath(path)
        mime_type, _ = mimetypes.guess_type(str(path_obj))
        fmt = (
            mime_type.removeprefix("audio/")
            if mime_type and mime_type.startswith("audio/")
            else "mp3"
        )

        return cls(data=base64.b64encode(path_obj.read_bytes()).decode(), format=fmt)

data instance-attribute

data: str

Audio data in base64 format.

format class-attribute instance-attribute

format: str | None = None

Audio format.

type class-attribute instance-attribute

type: Literal['audio_base64'] = Field('audio_base64', init=False)

Base64-encoded audio.

from_bytes classmethod

from_bytes(data: bytes, audio_format: str = 'mp3') -> Self

Create from raw bytes.

Source code in src/llmling_agent/models/content.py

@classmethod
def from_bytes(cls, data: bytes, audio_format: str = "mp3") -> Self:
    """Create from raw bytes."""
    return cls(data=base64.b64encode(data).decode(), format=audio_format)

from_path classmethod

from_path(path: StrPath) -> Self

Create from file path with auto format detection.

Source code in src/llmling_agent/models/content.py

@classmethod
def from_path(cls, path: StrPath) -> Self:
    """Create from file path with auto format detection."""
    import mimetypes

    from upath import UPath

    path_obj = UPath(path)
    mime_type, _ = mimetypes.guess_type(str(path_obj))
    fmt = (
        mime_type.removeprefix("audio/")
        if mime_type and mime_type.startswith("audio/")
        else "mp3"
    )

    return cls(data=base64.b64encode(path_obj.read_bytes()).decode(), format=fmt)

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for audio models.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for audio models."""
    data_url = f"data:audio/{self.format or 'mp3'};base64,{self.data}"
    content = {"url": data_url, "format": self.format or "auto"}
    return {"type": "audio", "audio": content}

AudioURLContent

Bases: AudioContent

Audio from URL.

Source code in src/llmling_agent/models/content.py

class AudioURLContent(AudioContent):
    """Audio from URL."""

    type: Literal["audio_url"] = Field("audio_url", init=False)
    """URL-based audio."""

    url: str
    """URL to the audio."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for audio models."""
        content = {"url": self.url, "format": self.format or "auto"}
        return {"type": "audio", "audio": content}

type class-attribute instance-attribute

type: Literal['audio_url'] = Field('audio_url', init=False)

URL-based audio.

url instance-attribute

url: str

URL to the audio.

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for audio models.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for audio models."""
    content = {"url": self.url, "format": self.format or "auto"}
    return {"type": "audio", "audio": content}

BaseTeam

Bases: MessageNode[TDeps, TResult]

Base class for Team and TeamRun.

Source code in src/llmling_agent/delegation/base_team.py

class BaseTeam[TDeps, TResult](MessageNode[TDeps, TResult]):
    """Base class for Team and TeamRun."""

    def __init__(
        self,
        agents: Sequence[MessageNode[TDeps, TResult]],
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        mcp_servers: list[str | MCPServerConfig] | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
    ):
        """Common variables only for typing."""
        from llmling_agent.delegation.teamrun import ExtendedTeamTalk

        self._name = name or " & ".join([i.name for i in agents])
        self.agents = EventedList[MessageNode]()
        self.agents.events.inserted.connect(self._on_node_added)
        self.agents.events.removed.connect(self._on_node_removed)
        self.agents.events.changed.connect(self._on_node_changed)
        super().__init__(
            name=self._name,
            context=self.context,
            mcp_servers=mcp_servers,
            description=description,
        )
        self.agents.extend(list(agents))
        self._team_talk = ExtendedTeamTalk()
        self.shared_prompt = shared_prompt
        self._main_task: asyncio.Task[Any] | None = None
        self._infinite = False
        self.picker = picker
        self.num_picks = num_picks
        self.pick_prompt = pick_prompt

    def to_tool(self, *, name: str | None = None, description: str | None = None) -> Tool:
        """Create a tool from this agent.

        Args:
            name: Optional tool name override
            description: Optional tool description override
        """
        tool_name = name or f"ask_{self.name}"

        async def wrapped_tool(prompt: str) -> TResult:
            result = await self.run(prompt)
            return result.data

        docstring = description or f"Get expert answer from node {self.name}"
        if self.description:
            docstring = f"{docstring}\n\n{self.description}"

        wrapped_tool.__doc__ = docstring
        wrapped_tool.__name__ = tool_name

        return Tool.from_callable(
            wrapped_tool,
            name_override=tool_name,
            description_override=docstring,
        )

    async def pick_agents(self, task: str) -> Sequence[MessageNode[Any, Any]]:
        """Pick agents to run."""
        if self.picker:
            if self.num_picks == 1:
                result = await self.picker.talk.pick(self, task, self.pick_prompt)
                return [result.selection]
            result = await self.picker.talk.pick_multiple(
                self,
                task,
                min_picks=self.num_picks or 1,
                max_picks=self.num_picks,
                prompt=self.pick_prompt,
            )
            return result.selections
        return list(self.agents)

    def _on_node_changed(self, index: int, old: MessageNode, new: MessageNode):
        """Handle node replacement in the agents list."""
        self._on_node_removed(index, old)
        self._on_node_added(index, new)

    def _on_node_added(self, index: int, node: MessageNode[Any, Any]):
        """Handler for adding nodes to the team."""
        from llmling_agent.agent import Agent, StructuredAgent

        if isinstance(node, Agent | StructuredAgent):
            node.tools.add_provider(self.mcp)
        # TODO: Right now connecting here is not desired since emission means db logging
        # ideally db logging would not rely on the "public" agent signal.

        # node.tool_used.connect(self.tool_used)

    def _on_node_removed(self, index: int, node: MessageNode[Any, Any]):
        """Handler for removing nodes from the team."""
        from llmling_agent.agent import Agent, StructuredAgent

        if isinstance(node, Agent | StructuredAgent):
            node.tools.remove_provider(self.mcp)
        # node.tool_used.disconnect(self.tool_used)

    def __repr__(self) -> str:
        """Create readable representation."""
        members = ", ".join(agent.name for agent in self.agents)
        name = f" ({self.name})" if self.name else ""
        return f"{self.__class__.__name__}[{len(self.agents)}]{name}: {members}"

    def __len__(self) -> int:
        """Get number of team members."""
        return len(self.agents)

    def __iter__(self) -> Iterator[MessageNode[TDeps, TResult]]:
        """Iterate over team members."""
        return iter(self.agents)

    def __getitem__(self, index_or_name: int | str) -> MessageNode[TDeps, TResult]:
        """Get team member by index or name."""
        if isinstance(index_or_name, str):
            return next(agent for agent in self.agents if agent.name == index_or_name)
        return self.agents[index_or_name]

    def __or__(
        self,
        other: AnyAgent[Any, Any] | ProcessorCallback[Any] | BaseTeam[Any, Any],
    ) -> TeamRun[Any, Any]:
        """Create a sequential pipeline."""
        from llmling_agent.agent import Agent, StructuredAgent
        from llmling_agent.delegation.teamrun import TeamRun

        # Handle conversion of callables first
        if callable(other):
            if has_return_type(other, str):
                other = Agent.from_callback(other)
            else:
                other = StructuredAgent.from_callback(other)
            other.context.pool = self.context.pool

        # If we're already a TeamRun, extend it
        if isinstance(self, TeamRun):
            if self.validator:
                # If we have a validator, create new TeamRun to preserve validation
                return TeamRun([self, other])
            self.agents.append(other)
            return self
        # Otherwise create new TeamRun
        return TeamRun([self, other])

    @overload
    def __and__(self, other: Team[None]) -> Team[None]: ...

    @overload
    def __and__(self, other: Team[TDeps]) -> Team[TDeps]: ...

    @overload
    def __and__(self, other: Team[Any]) -> Team[Any]: ...

    @overload
    def __and__(self, other: AnyAgent[TDeps, Any]) -> Team[TDeps]: ...

    @overload
    def __and__(self, other: AnyAgent[Any, Any]) -> Team[Any]: ...

    def __and__(
        self, other: Team[Any] | AnyAgent[Any, Any] | ProcessorCallback[Any]
    ) -> Team[Any]:
        """Combine teams, preserving type safety for same types."""
        from llmling_agent.agent import Agent, StructuredAgent
        from llmling_agent.delegation.team import Team

        if callable(other):
            if has_return_type(other, str):
                other = Agent.from_callback(other)
            else:
                other = StructuredAgent.from_callback(other)
            other.context.pool = self.context.pool

        match other:
            case Team():
                # Flatten when combining Teams
                return Team([*self.agents, *other.agents])
            case _:
                # Everything else just becomes a member
                return Team([*self.agents, other])

    @property
    def stats(self) -> AggregatedMessageStats:
        """Get aggregated stats from all team members."""
        return AggregatedMessageStats(stats=[agent.stats for agent in self.agents])

    @property
    def is_running(self) -> bool:
        """Whether execution is currently running."""
        return bool(self._main_task and not self._main_task.done())

    def is_busy(self) -> bool:
        """Check if team is processing any tasks."""
        return bool(self._pending_tasks or self._main_task)

    async def stop(self):
        """Stop background execution if running."""
        if self._main_task and not self._main_task.done():
            self._main_task.cancel()
            await self._main_task
        self._main_task = None
        await self.cleanup_tasks()

    async def wait(self) -> ChatMessage[Any] | None:
        """Wait for background execution to complete and return last message."""
        if not self._main_task:
            msg = "No execution running"
            raise RuntimeError(msg)
        if self._infinite:
            msg = "Cannot wait on infinite execution"
            raise RuntimeError(msg)
        try:
            return await self._main_task
        finally:
            await self.cleanup_tasks()
            self._main_task = None

    async def run_in_background(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        max_count: int | None = 1,  # 1 = single execution, None = indefinite
        interval: float = 1.0,
        **kwargs: Any,
    ) -> ExtendedTeamTalk:
        """Start execution in background.

        Args:
            prompts: Prompts to execute
            max_count: Maximum number of executions (None = run indefinitely)
            interval: Seconds between executions
            **kwargs: Additional args for execute()
        """
        if self._main_task:
            msg = "Execution already running"
            raise RuntimeError(msg)
        self._infinite = max_count is None

        async def _continuous() -> ChatMessage[Any] | None:
            count = 0
            last_message = None
            while max_count is None or count < max_count:
                try:
                    result = await self.execute(*prompts, **kwargs)
                    last_message = result[-1].message if result else None
                    count += 1
                    if max_count is None or count < max_count:
                        await asyncio.sleep(interval)
                except asyncio.CancelledError:
                    logger.debug("Background execution cancelled")
                    break
            return last_message

        self._main_task = self.create_task(_continuous(), name="main_execution")
        return self._team_talk

    @property
    def execution_stats(self) -> AggregatedTalkStats:
        """Get current execution statistics."""
        return self._team_talk.stats

    @property
    def talk(self) -> ExtendedTeamTalk:
        """Get current connection."""
        return self._team_talk

    @property
    def events(self) -> ListEvents:
        """Get events for the team."""
        return self.agents.events

    async def cancel(self):
        """Cancel execution and cleanup."""
        if self._main_task:
            self._main_task.cancel()
        await self.cleanup_tasks()

    def get_structure_diagram(self) -> str:
        """Generate mermaid flowchart of node hierarchy."""
        lines = ["flowchart TD"]

        def add_node(node: MessageNode[Any, Any], parent: str | None = None):
            """Recursively add node and its members to diagram."""
            node_id = f"node_{id(node)}"
            lines.append(f"    {node_id}[{node.name}]")
            if parent:
                lines.append(f"    {parent} --> {node_id}")

            # If it's a team, recursively add its members
            from llmling_agent.delegation.base_team import BaseTeam

            if isinstance(node, BaseTeam):
                for member in node.agents:
                    add_node(member, node_id)

        # Start with root nodes (team members)
        for node in self.agents:
            add_node(node)

        return "\n".join(lines)

    def iter_agents(self) -> Iterator[AnyAgent[Any, Any]]:
        """Recursively iterate over all child agents."""
        from llmling_agent.agent import Agent, StructuredAgent

        for node in self.agents:
            match node:
                case BaseTeam():
                    yield from node.iter_agents()
                case Agent() | StructuredAgent():
                    yield node
                case _:
                    msg = f"Invalid node type: {type(node)}"
                    raise ValueError(msg)

    @property
    def context(self) -> TeamContext:
        """Get shared pool from team members.

        Raises:
            ValueError: If team members belong to different pools
        """
        from llmling_agent.delegation.team import Team

        pool_ids: set[int] = set()
        shared_pool: AgentPool | None = None
        team_config: TeamConfig | None = None

        for agent in self.iter_agents():
            if agent.context and agent.context.pool:
                pool_id = id(agent.context.pool)
                if pool_id not in pool_ids:
                    pool_ids.add(pool_id)
                    shared_pool = agent.context.pool
                    if shared_pool.manifest.teams:
                        team_config = shared_pool.manifest.teams.get(self.name)
        if not team_config:
            mode = "parallel" if isinstance(self, Team) else "sequential"
            team_config = TeamConfig(name=self.name, mode=mode, members=[])
        if not pool_ids:
            logger.info("No pool found for team %s.", self.name)
            return TeamContext(
                node_name=self.name,
                pool=shared_pool,
                config=team_config,
                definition=shared_pool.manifest if shared_pool else AgentsManifest(),
            )

        if len(pool_ids) > 1:
            msg = f"Team members in {self.name} belong to different pools"
            raise ValueError(msg)
        return TeamContext(
            node_name=self.name,
            pool=shared_pool,
            config=team_config,
            definition=shared_pool.manifest if shared_pool else AgentsManifest(),
        )

    @context.setter
    def context(self, value: NodeContext):
        msg = "Cannot set context on BaseTeam"
        raise RuntimeError(msg)

    async def distribute(
        self,
        content: str,
        *,
        tools: list[str] | None = None,
        resources: list[str] | None = None,
        metadata: dict[str, Any] | None = None,
    ):
        """Distribute content and capabilities to all team members."""
        for agent in self.iter_agents():
            # Add context message
            agent.conversation.add_context_message(
                content, source="distribution", metadata=metadata
            )

            # Register tools if provided
            if tools:
                for tool in tools:
                    agent.tools.register_tool(tool)

            # Load resources if provided
            if resources:
                for resource in resources:
                    await agent.conversation.load_context_source(resource)

    @asynccontextmanager
    async def temporary_state(
        self,
        *,
        system_prompts: list[AnyPromptType] | None = None,
        replace_prompts: bool = False,
        tools: list[ToolType] | None = None,
        replace_tools: bool = False,
        history: list[AnyPromptType] | SessionQuery | None = None,
        replace_history: bool = False,
        pause_routing: bool = False,
        model: ModelType | None = None,
        provider: AgentProvider | None = None,
    ) -> AsyncIterator[Self]:
        """Temporarily modify state of all agents in the team.

        All agents in the team will enter their temporary state simultaneously.

        Args:
            system_prompts: Temporary system prompts to use
            replace_prompts: Whether to replace existing prompts
            tools: Temporary tools to make available
            replace_tools: Whether to replace existing tools
            history: Conversation history (prompts or query)
            replace_history: Whether to replace existing history
            pause_routing: Whether to pause message routing
            model: Temporary model override
            provider: Temporary provider override
        """
        # Get all agents (flattened) before entering context
        agents = list(self.iter_agents())

        async with AsyncExitStack() as stack:
            if pause_routing:
                await stack.enter_async_context(self.connections.paused_routing())
            # Enter temporary state for all agents
            for agent in agents:
                await stack.enter_async_context(
                    agent.temporary_state(
                        system_prompts=system_prompts,
                        replace_prompts=replace_prompts,
                        tools=tools,
                        replace_tools=replace_tools,
                        history=history,
                        replace_history=replace_history,
                        pause_routing=pause_routing,
                        model=model,
                        provider=provider,
                    )
                )
            try:
                yield self
            finally:
                # AsyncExitStack will handle cleanup of all states
                pass

    @abstractmethod
    async def execute(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        **kwargs: Any,
    ) -> TeamResponse: ...

    def run_sync(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        store_history: bool = True,
    ) -> ChatMessage[TResult]:
        """Run agent synchronously (convenience wrapper).

        Args:
            prompt: User query or instruction
            store_history: Whether the message exchange should be added to the
                           context window
        Returns:
            Result containing response and run information
        """
        coro = self.run(*prompt, store_history=store_history)
        return self.run_task_sync(coro)

context property writable

context: TeamContext

Get shared pool from team members.

Raises:

Type	Description
ValueError	If team members belong to different pools

events property

events: ListEvents

Get events for the team.

execution_stats property

execution_stats: AggregatedTalkStats

Get current execution statistics.

is_running property

is_running: bool

Whether execution is currently running.

stats property

stats: AggregatedMessageStats

Get aggregated stats from all team members.

talk property

talk: ExtendedTeamTalk

Get current connection.

__and__

__and__(other: Team[None]) -> Team[None]

__and__(other: Team[TDeps]) -> Team[TDeps]

__and__(other: Team[Any]) -> Team[Any]

__and__(other: AnyAgent[TDeps, Any]) -> Team[TDeps]

__and__(other: AnyAgent[Any, Any]) -> Team[Any]

__and__(other: Team[Any] | AnyAgent[Any, Any] | ProcessorCallback[Any]) -> Team[Any]

Combine teams, preserving type safety for same types.

Source code in src/llmling_agent/delegation/base_team.py

def __and__(
    self, other: Team[Any] | AnyAgent[Any, Any] | ProcessorCallback[Any]
) -> Team[Any]:
    """Combine teams, preserving type safety for same types."""
    from llmling_agent.agent import Agent, StructuredAgent
    from llmling_agent.delegation.team import Team

    if callable(other):
        if has_return_type(other, str):
            other = Agent.from_callback(other)
        else:
            other = StructuredAgent.from_callback(other)
        other.context.pool = self.context.pool

    match other:
        case Team():
            # Flatten when combining Teams
            return Team([*self.agents, *other.agents])
        case _:
            # Everything else just becomes a member
            return Team([*self.agents, other])

__getitem__

__getitem__(index_or_name: int | str) -> MessageNode[TDeps, TResult]

Get team member by index or name.

Source code in src/llmling_agent/delegation/base_team.py

def __getitem__(self, index_or_name: int | str) -> MessageNode[TDeps, TResult]:
    """Get team member by index or name."""
    if isinstance(index_or_name, str):
        return next(agent for agent in self.agents if agent.name == index_or_name)
    return self.agents[index_or_name]

__init__

__init__(
    agents: Sequence[MessageNode[TDeps, TResult]],
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    mcp_servers: list[str | MCPServerConfig] | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
)

Common variables only for typing.

Source code in src/llmling_agent/delegation/base_team.py

def __init__(
    self,
    agents: Sequence[MessageNode[TDeps, TResult]],
    *,
    name: str | None = None,
    description: str | None = None,
    shared_prompt: str | None = None,
    mcp_servers: list[str | MCPServerConfig] | None = None,
    picker: AnyAgent[Any, Any] | None = None,
    num_picks: int | None = None,
    pick_prompt: str | None = None,
):
    """Common variables only for typing."""
    from llmling_agent.delegation.teamrun import ExtendedTeamTalk

    self._name = name or " & ".join([i.name for i in agents])
    self.agents = EventedList[MessageNode]()
    self.agents.events.inserted.connect(self._on_node_added)
    self.agents.events.removed.connect(self._on_node_removed)
    self.agents.events.changed.connect(self._on_node_changed)
    super().__init__(
        name=self._name,
        context=self.context,
        mcp_servers=mcp_servers,
        description=description,
    )
    self.agents.extend(list(agents))
    self._team_talk = ExtendedTeamTalk()
    self.shared_prompt = shared_prompt
    self._main_task: asyncio.Task[Any] | None = None
    self._infinite = False
    self.picker = picker
    self.num_picks = num_picks
    self.pick_prompt = pick_prompt

__iter__

__iter__() -> Iterator[MessageNode[TDeps, TResult]]

Iterate over team members.

Source code in src/llmling_agent/delegation/base_team.py

def __iter__(self) -> Iterator[MessageNode[TDeps, TResult]]:
    """Iterate over team members."""
    return iter(self.agents)

__len__

__len__() -> int

Get number of team members.

Source code in src/llmling_agent/delegation/base_team.py

def __len__(self) -> int:
    """Get number of team members."""
    return len(self.agents)

__or__

__or__(
    other: AnyAgent[Any, Any] | ProcessorCallback[Any] | BaseTeam[Any, Any],
) -> TeamRun[Any, Any]

Create a sequential pipeline.

Source code in src/llmling_agent/delegation/base_team.py

def __or__(
    self,
    other: AnyAgent[Any, Any] | ProcessorCallback[Any] | BaseTeam[Any, Any],
) -> TeamRun[Any, Any]:
    """Create a sequential pipeline."""
    from llmling_agent.agent import Agent, StructuredAgent
    from llmling_agent.delegation.teamrun import TeamRun

    # Handle conversion of callables first
    if callable(other):
        if has_return_type(other, str):
            other = Agent.from_callback(other)
        else:
            other = StructuredAgent.from_callback(other)
        other.context.pool = self.context.pool

    # If we're already a TeamRun, extend it
    if isinstance(self, TeamRun):
        if self.validator:
            # If we have a validator, create new TeamRun to preserve validation
            return TeamRun([self, other])
        self.agents.append(other)
        return self
    # Otherwise create new TeamRun
    return TeamRun([self, other])

__repr__

__repr__() -> str

Create readable representation.

Source code in src/llmling_agent/delegation/base_team.py

def __repr__(self) -> str:
    """Create readable representation."""
    members = ", ".join(agent.name for agent in self.agents)
    name = f" ({self.name})" if self.name else ""
    return f"{self.__class__.__name__}[{len(self.agents)}]{name}: {members}"

cancel async

cancel()

Cancel execution and cleanup.

Source code in src/llmling_agent/delegation/base_team.py

async def cancel(self):
    """Cancel execution and cleanup."""
    if self._main_task:
        self._main_task.cancel()
    await self.cleanup_tasks()

distribute async

distribute(
    content: str,
    *,
    tools: list[str] | None = None,
    resources: list[str] | None = None,
    metadata: dict[str, Any] | None = None,
)

Distribute content and capabilities to all team members.

Source code in src/llmling_agent/delegation/base_team.py

async def distribute(
    self,
    content: str,
    *,
    tools: list[str] | None = None,
    resources: list[str] | None = None,
    metadata: dict[str, Any] | None = None,
):
    """Distribute content and capabilities to all team members."""
    for agent in self.iter_agents():
        # Add context message
        agent.conversation.add_context_message(
            content, source="distribution", metadata=metadata
        )

        # Register tools if provided
        if tools:
            for tool in tools:
                agent.tools.register_tool(tool)

        # Load resources if provided
        if resources:
            for resource in resources:
                await agent.conversation.load_context_source(resource)

get_structure_diagram

get_structure_diagram() -> str

Generate mermaid flowchart of node hierarchy.

Source code in src/llmling_agent/delegation/base_team.py

def get_structure_diagram(self) -> str:
    """Generate mermaid flowchart of node hierarchy."""
    lines = ["flowchart TD"]

    def add_node(node: MessageNode[Any, Any], parent: str | None = None):
        """Recursively add node and its members to diagram."""
        node_id = f"node_{id(node)}"
        lines.append(f"    {node_id}[{node.name}]")
        if parent:
            lines.append(f"    {parent} --> {node_id}")

        # If it's a team, recursively add its members
        from llmling_agent.delegation.base_team import BaseTeam

        if isinstance(node, BaseTeam):
            for member in node.agents:
                add_node(member, node_id)

    # Start with root nodes (team members)
    for node in self.agents:
        add_node(node)

    return "\n".join(lines)

is_busy

is_busy() -> bool

Check if team is processing any tasks.

Source code in src/llmling_agent/delegation/base_team.py

def is_busy(self) -> bool:
    """Check if team is processing any tasks."""
    return bool(self._pending_tasks or self._main_task)

iter_agents

iter_agents() -> Iterator[AnyAgent[Any, Any]]

Recursively iterate over all child agents.

Source code in src/llmling_agent/delegation/base_team.py

def iter_agents(self) -> Iterator[AnyAgent[Any, Any]]:
    """Recursively iterate over all child agents."""
    from llmling_agent.agent import Agent, StructuredAgent

    for node in self.agents:
        match node:
            case BaseTeam():
                yield from node.iter_agents()
            case Agent() | StructuredAgent():
                yield node
            case _:
                msg = f"Invalid node type: {type(node)}"
                raise ValueError(msg)

pick_agents async

pick_agents(task: str) -> Sequence[MessageNode[Any, Any]]

Pick agents to run.

Source code in src/llmling_agent/delegation/base_team.py

async def pick_agents(self, task: str) -> Sequence[MessageNode[Any, Any]]:
    """Pick agents to run."""
    if self.picker:
        if self.num_picks == 1:
            result = await self.picker.talk.pick(self, task, self.pick_prompt)
            return [result.selection]
        result = await self.picker.talk.pick_multiple(
            self,
            task,
            min_picks=self.num_picks or 1,
            max_picks=self.num_picks,
            prompt=self.pick_prompt,
        )
        return result.selections
    return list(self.agents)

run_in_background async

run_in_background(
    *prompts: AnyPromptType | Image | PathLike[str] | None,
    max_count: int | None = 1,
    interval: float = 1.0,
    **kwargs: Any,
) -> ExtendedTeamTalk

Start execution in background.

Parameters:

Name	Type	Description	Default
prompts	AnyPromptType | Image | PathLike[str] | None	Prompts to execute	()
max_count	int | None	Maximum number of executions (None = run indefinitely)	1
interval	float	Seconds between executions	1.0
**kwargs	Any	Additional args for execute()	{}

Source code in src/llmling_agent/delegation/base_team.py

async def run_in_background(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
    max_count: int | None = 1,  # 1 = single execution, None = indefinite
    interval: float = 1.0,
    **kwargs: Any,
) -> ExtendedTeamTalk:
    """Start execution in background.

    Args:
        prompts: Prompts to execute
        max_count: Maximum number of executions (None = run indefinitely)
        interval: Seconds between executions
        **kwargs: Additional args for execute()
    """
    if self._main_task:
        msg = "Execution already running"
        raise RuntimeError(msg)
    self._infinite = max_count is None

    async def _continuous() -> ChatMessage[Any] | None:
        count = 0
        last_message = None
        while max_count is None or count < max_count:
            try:
                result = await self.execute(*prompts, **kwargs)
                last_message = result[-1].message if result else None
                count += 1
                if max_count is None or count < max_count:
                    await asyncio.sleep(interval)
            except asyncio.CancelledError:
                logger.debug("Background execution cancelled")
                break
        return last_message

    self._main_task = self.create_task(_continuous(), name="main_execution")
    return self._team_talk

run_sync

run_sync(
    *prompt: AnyPromptType | Image | PathLike[str], store_history: bool = True
) -> ChatMessage[TResult]

Run agent synchronously (convenience wrapper).

Parameters:

Name	Type	Description	Default
prompt	AnyPromptType | Image | PathLike[str]	User query or instruction	()
store_history	bool	Whether the message exchange should be added to the context window	True

Returns: Result containing response and run information

Source code in src/llmling_agent/delegation/base_team.py

def run_sync(
    self,
    *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    store_history: bool = True,
) -> ChatMessage[TResult]:
    """Run agent synchronously (convenience wrapper).

    Args:
        prompt: User query or instruction
        store_history: Whether the message exchange should be added to the
                       context window
    Returns:
        Result containing response and run information
    """
    coro = self.run(*prompt, store_history=store_history)
    return self.run_task_sync(coro)

stop async

stop()

Stop background execution if running.

Source code in src/llmling_agent/delegation/base_team.py

async def stop(self):
    """Stop background execution if running."""
    if self._main_task and not self._main_task.done():
        self._main_task.cancel()
        await self._main_task
    self._main_task = None
    await self.cleanup_tasks()

temporary_state async

temporary_state(
    *,
    system_prompts: list[AnyPromptType] | None = None,
    replace_prompts: bool = False,
    tools: list[ToolType] | None = None,
    replace_tools: bool = False,
    history: list[AnyPromptType] | SessionQuery | None = None,
    replace_history: bool = False,
    pause_routing: bool = False,
    model: ModelType | None = None,
    provider: AgentProvider | None = None,
) -> AsyncIterator[Self]

Temporarily modify state of all agents in the team.

All agents in the team will enter their temporary state simultaneously.

Parameters:

Name	Type	Description	Default
system_prompts	list[AnyPromptType] | None	Temporary system prompts to use	None
replace_prompts	bool	Whether to replace existing prompts	False
tools	list[ToolType] | None	Temporary tools to make available	None
replace_tools	bool	Whether to replace existing tools	False
history	list[AnyPromptType] | SessionQuery | None	Conversation history (prompts or query)	None
replace_history	bool	Whether to replace existing history	False
pause_routing	bool	Whether to pause message routing	False
model	ModelType | None	Temporary model override	None
provider	AgentProvider | None	Temporary provider override	None

Source code in src/llmling_agent/delegation/base_team.py

@asynccontextmanager
async def temporary_state(
    self,
    *,
    system_prompts: list[AnyPromptType] | None = None,
    replace_prompts: bool = False,
    tools: list[ToolType] | None = None,
    replace_tools: bool = False,
    history: list[AnyPromptType] | SessionQuery | None = None,
    replace_history: bool = False,
    pause_routing: bool = False,
    model: ModelType | None = None,
    provider: AgentProvider | None = None,
) -> AsyncIterator[Self]:
    """Temporarily modify state of all agents in the team.

    All agents in the team will enter their temporary state simultaneously.

    Args:
        system_prompts: Temporary system prompts to use
        replace_prompts: Whether to replace existing prompts
        tools: Temporary tools to make available
        replace_tools: Whether to replace existing tools
        history: Conversation history (prompts or query)
        replace_history: Whether to replace existing history
        pause_routing: Whether to pause message routing
        model: Temporary model override
        provider: Temporary provider override
    """
    # Get all agents (flattened) before entering context
    agents = list(self.iter_agents())

    async with AsyncExitStack() as stack:
        if pause_routing:
            await stack.enter_async_context(self.connections.paused_routing())
        # Enter temporary state for all agents
        for agent in agents:
            await stack.enter_async_context(
                agent.temporary_state(
                    system_prompts=system_prompts,
                    replace_prompts=replace_prompts,
                    tools=tools,
                    replace_tools=replace_tools,
                    history=history,
                    replace_history=replace_history,
                    pause_routing=pause_routing,
                    model=model,
                    provider=provider,
                )
            )
        try:
            yield self
        finally:
            # AsyncExitStack will handle cleanup of all states
            pass

to_tool

to_tool(*, name: str | None = None, description: str | None = None) -> Tool

Create a tool from this agent.

Parameters:

Name	Type	Description	Default
name	str | None	Optional tool name override	None
description	str | None	Optional tool description override	None

Source code in src/llmling_agent/delegation/base_team.py

def to_tool(self, *, name: str | None = None, description: str | None = None) -> Tool:
    """Create a tool from this agent.

    Args:
        name: Optional tool name override
        description: Optional tool description override
    """
    tool_name = name or f"ask_{self.name}"

    async def wrapped_tool(prompt: str) -> TResult:
        result = await self.run(prompt)
        return result.data

    docstring = description or f"Get expert answer from node {self.name}"
    if self.description:
        docstring = f"{docstring}\n\n{self.description}"

    wrapped_tool.__doc__ = docstring
    wrapped_tool.__name__ = tool_name

    return Tool.from_callable(
        wrapped_tool,
        name_override=tool_name,
        description_override=docstring,
    )

wait async

wait() -> ChatMessage[Any] | None

Wait for background execution to complete and return last message.

Source code in src/llmling_agent/delegation/base_team.py

async def wait(self) -> ChatMessage[Any] | None:
    """Wait for background execution to complete and return last message."""
    if not self._main_task:
        msg = "No execution running"
        raise RuntimeError(msg)
    if self._infinite:
        msg = "Cannot wait on infinite execution"
        raise RuntimeError(msg)
    try:
        return await self._main_task
    finally:
        await self.cleanup_tasks()
        self._main_task = None

ChatMessage dataclass

Common message format for all UI types.

Generically typed with: ChatMessage[Type of Content] The type can either be str or a BaseModel subclass.

Source code in src/llmling_agent/messaging/messages.py

@dataclass
class ChatMessage[TContent]:
    """Common message format for all UI types.

    Generically typed with: ChatMessage[Type of Content]
    The type can either be str or a BaseModel subclass.
    """

    content: TContent
    """Message content, typed as TContent (either str or BaseModel)."""

    role: MessageRole
    """Role of the message sender (user/assistant/system)."""

    model: str | None = None
    """Name of the model that generated this message."""

    metadata: SimpleJsonType = field(default_factory=dict)
    """Additional metadata about the message."""

    timestamp: datetime = field(default_factory=get_now)
    """When this message was created."""

    cost_info: TokenCost | None = None
    """Token usage and costs for this specific message if available."""

    message_id: str = field(default_factory=lambda: str(uuid4()))
    """Unique identifier for this message."""

    conversation_id: str | None = None
    """ID of the conversation this message belongs to."""

    response_time: float | None = None
    """Time it took the LLM to respond."""

    tool_calls: list[ToolCallInfo] = field(default_factory=list)
    """List of tool calls made during message generation."""

    associated_messages: list[ChatMessage[Any]] = field(default_factory=list)
    """List of messages which were generated during the the creation of this messsage."""

    name: str | None = None
    """Display name for the message sender in UI."""

    forwarded_from: list[str] = field(default_factory=list)
    """List of agent names (the chain) that forwarded this message to the sender."""

    provider_extra: dict[str, Any] = field(default_factory=dict)
    """Provider specific metadata / extra information."""

    @classmethod
    def from_openai_format(
        cls,
        message: dict[str, Any],
        conversation_id: str | None = None,
    ) -> ChatMessage[str]:
        """Create ChatMessage from OpenAI message format.

        Args:
            message: OpenAI format message dict with role, content etc.
            conversation_id: Optional conversation ID to assign

        Returns:
            Converted ChatMessage

        Example:
            >>> msg = ChatMessage.from_openai_format({
            ...     "role": "user",
            ...     "content": "Hello!",
            ...     "name": "john"
            ... })
        """
        # Handle multimodal content lists (OpenAI vision format)
        if isinstance(message["content"], list):
            # Combine text parts
            content = "\n".join(
                part["text"] for part in message["content"] if part["type"] == "text"
            )
        else:
            content = message["content"] or ""

        return ChatMessage[str](
            content=str(content),
            role=message["role"],
            name=message.get("name"),
            conversation_id=conversation_id,
            tool_calls=[
                ToolCallInfo(
                    agent_name=message.get("name") or "",
                    tool_call_id=tc["id"],
                    tool_name=tc["function"]["name"],
                    args=tc["function"]["arguments"],
                    result=None,  # OpenAI format doesn't include results
                )
                for tc in message.get("tool_calls", [])
            ]
            if message.get("tool_calls")
            else [],
            metadata={"function_call": message["function_call"]}
            if "function_call" in message
            else {},
        )

    def forwarded(self, previous_message: ChatMessage[Any]) -> Self:
        """Create new message showing it was forwarded from another message.

        Args:
            previous_message: The message that led to this one's creation

        Returns:
            New message with updated chain showing the path through previous message
        """
        from_ = [*previous_message.forwarded_from, previous_message.name or "unknown"]
        return replace(self, forwarded_from=from_)

    def to_text_message(self) -> ChatMessage[str]:
        """Convert this message to a text-only version."""
        return dataclasses.replace(self, content=str(self.content))  # type: ignore

    def _get_content_str(self) -> str:
        """Get string representation of content."""
        match self.content:
            case str():
                return self.content
            case BaseModel():
                return self.content.model_dump_json(indent=2)
            case _:
                msg = f"Unexpected content type: {type(self.content)}"
                raise ValueError(msg)

    @property
    def data(self) -> TContent:
        """Get content as typed data. Provides compat to AgentRunResult."""
        return self.content

    def format(
        self,
        style: FormatStyle = "simple",
        *,
        template: str | None = None,
        variables: dict[str, Any] | None = None,
        show_metadata: bool = False,
        show_costs: bool = False,
    ) -> str:
        """Format message with configurable style.

        Args:
            style: Predefined style or "custom" for custom template
            template: Custom Jinja template (required if style="custom")
            variables: Additional variables for template rendering
            show_metadata: Whether to include metadata
            show_costs: Whether to include cost information

        Raises:
            ValueError: If style is "custom" but no template provided
                    or if style is invalid
        """
        from jinjarope import Environment
        import yamling

        env = Environment(trim_blocks=True, lstrip_blocks=True)
        env.filters["to_yaml"] = yamling.dump_yaml

        match style:
            case "custom":
                if not template:
                    msg = "Custom style requires a template"
                    raise ValueError(msg)
                template_str = template
            case _ if style in MESSAGE_TEMPLATES:
                template_str = MESSAGE_TEMPLATES[style]
            case _:
                msg = f"Invalid style: {style}"
                raise ValueError(msg)

        template_obj = env.from_string(template_str)
        vars_ = {**asdict(self), "show_metadata": show_metadata, "show_costs": show_costs}
        if variables:
            vars_.update(variables)

        return template_obj.render(**vars_)

associated_messages class-attribute instance-attribute

associated_messages: list[ChatMessage[Any]] = field(default_factory=list)

List of messages which were generated during the the creation of this messsage.

content instance-attribute

content: TContent

Message content, typed as TContent (either str or BaseModel).

conversation_id class-attribute instance-attribute

conversation_id: str | None = None

ID of the conversation this message belongs to.

cost_info class-attribute instance-attribute

cost_info: TokenCost | None = None

Token usage and costs for this specific message if available.

data property

data: TContent

Get content as typed data. Provides compat to AgentRunResult.

forwarded_from class-attribute instance-attribute

forwarded_from: list[str] = field(default_factory=list)

List of agent names (the chain) that forwarded this message to the sender.

message_id class-attribute instance-attribute

message_id: str = field(default_factory=lambda: str(uuid4()))

Unique identifier for this message.

metadata class-attribute instance-attribute

metadata: SimpleJsonType = field(default_factory=dict)

Additional metadata about the message.

model class-attribute instance-attribute

model: str | None = None

Name of the model that generated this message.

name class-attribute instance-attribute

name: str | None = None

Display name for the message sender in UI.

provider_extra class-attribute instance-attribute

provider_extra: dict[str, Any] = field(default_factory=dict)

Provider specific metadata / extra information.

response_time class-attribute instance-attribute

response_time: float | None = None

Time it took the LLM to respond.

role instance-attribute

role: MessageRole

Role of the message sender (user/assistant/system).

timestamp class-attribute instance-attribute

timestamp: datetime = field(default_factory=get_now)

When this message was created.

tool_calls class-attribute instance-attribute

tool_calls: list[ToolCallInfo] = field(default_factory=list)

List of tool calls made during message generation.

format

format(
    style: FormatStyle = "simple",
    *,
    template: str | None = None,
    variables: dict[str, Any] | None = None,
    show_metadata: bool = False,
    show_costs: bool = False,
) -> str

Format message with configurable style.

Parameters:

Name	Type	Description	Default
style	FormatStyle	Predefined style or "custom" for custom template	'simple'
template	str | None	Custom Jinja template (required if style="custom")	None
variables	dict[str, Any] | None	Additional variables for template rendering	None
show_metadata	bool	Whether to include metadata	False
show_costs	bool	Whether to include cost information	False

Raises:

Type	Description
ValueError	If style is "custom" but no template provided or if style is invalid

Source code in src/llmling_agent/messaging/messages.py

def format(
    self,
    style: FormatStyle = "simple",
    *,
    template: str | None = None,
    variables: dict[str, Any] | None = None,
    show_metadata: bool = False,
    show_costs: bool = False,
) -> str:
    """Format message with configurable style.

    Args:
        style: Predefined style or "custom" for custom template
        template: Custom Jinja template (required if style="custom")
        variables: Additional variables for template rendering
        show_metadata: Whether to include metadata
        show_costs: Whether to include cost information

    Raises:
        ValueError: If style is "custom" but no template provided
                or if style is invalid
    """
    from jinjarope import Environment
    import yamling

    env = Environment(trim_blocks=True, lstrip_blocks=True)
    env.filters["to_yaml"] = yamling.dump_yaml

    match style:
        case "custom":
            if not template:
                msg = "Custom style requires a template"
                raise ValueError(msg)
            template_str = template
        case _ if style in MESSAGE_TEMPLATES:
            template_str = MESSAGE_TEMPLATES[style]
        case _:
            msg = f"Invalid style: {style}"
            raise ValueError(msg)

    template_obj = env.from_string(template_str)
    vars_ = {**asdict(self), "show_metadata": show_metadata, "show_costs": show_costs}
    if variables:
        vars_.update(variables)

    return template_obj.render(**vars_)

forwarded

forwarded(previous_message: ChatMessage[Any]) -> Self

Create new message showing it was forwarded from another message.

Parameters:

Name	Type	Description	Default
previous_message	ChatMessage[Any]	The message that led to this one's creation	required

Returns:

Type	Description
Self	New message with updated chain showing the path through previous message

Source code in src/llmling_agent/messaging/messages.py

def forwarded(self, previous_message: ChatMessage[Any]) -> Self:
    """Create new message showing it was forwarded from another message.

    Args:
        previous_message: The message that led to this one's creation

    Returns:
        New message with updated chain showing the path through previous message
    """
    from_ = [*previous_message.forwarded_from, previous_message.name or "unknown"]
    return replace(self, forwarded_from=from_)

from_openai_format classmethod

from_openai_format(
    message: dict[str, Any], conversation_id: str | None = None
) -> ChatMessage[str]

Create ChatMessage from OpenAI message format.

Parameters:

Name	Type	Description	Default
message	dict[str, Any]	OpenAI format message dict with role, content etc.	required
conversation_id	str | None	Optional conversation ID to assign	None

Returns:

Type	Description
ChatMessage[str]	Converted ChatMessage

Example

msg = ChatMessage.from_openai_format({ ... "role": "user", ... "content": "Hello!", ... "name": "john" ... })

Source code in src/llmling_agent/messaging/messages.py

@classmethod
def from_openai_format(
    cls,
    message: dict[str, Any],
    conversation_id: str | None = None,
) -> ChatMessage[str]:
    """Create ChatMessage from OpenAI message format.

    Args:
        message: OpenAI format message dict with role, content etc.
        conversation_id: Optional conversation ID to assign

    Returns:
        Converted ChatMessage

    Example:
        >>> msg = ChatMessage.from_openai_format({
        ...     "role": "user",
        ...     "content": "Hello!",
        ...     "name": "john"
        ... })
    """
    # Handle multimodal content lists (OpenAI vision format)
    if isinstance(message["content"], list):
        # Combine text parts
        content = "\n".join(
            part["text"] for part in message["content"] if part["type"] == "text"
        )
    else:
        content = message["content"] or ""

    return ChatMessage[str](
        content=str(content),
        role=message["role"],
        name=message.get("name"),
        conversation_id=conversation_id,
        tool_calls=[
            ToolCallInfo(
                agent_name=message.get("name") or "",
                tool_call_id=tc["id"],
                tool_name=tc["function"]["name"],
                args=tc["function"]["arguments"],
                result=None,  # OpenAI format doesn't include results
            )
            for tc in message.get("tool_calls", [])
        ]
        if message.get("tool_calls")
        else [],
        metadata={"function_call": message["function_call"]}
        if "function_call" in message
        else {},
    )

to_text_message

to_text_message() -> ChatMessage[str]

Convert this message to a text-only version.

Source code in src/llmling_agent/messaging/messages.py

def to_text_message(self) -> ChatMessage[str]:
    """Convert this message to a text-only version."""
    return dataclasses.replace(self, content=str(self.content))  # type: ignore

ImageBase64Content

Bases: BaseImageContent

Image from base64 data.

Source code in src/llmling_agent/models/content.py

class ImageBase64Content(BaseImageContent):
    """Image from base64 data."""

    type: Literal["image_base64"] = Field("image_base64", init=False)
    """Base64-encoded image."""

    data: str
    """Base64-encoded image data."""

    mime_type: str = "image/jpeg"
    """MIME type of the image."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for vision models."""
        data_url = f"data:{self.mime_type};base64,{self.data}"
        content = {"url": data_url, "detail": self.detail or "auto"}
        return {"type": "image_url", "image_url": content}

    @classmethod
    def from_bytes(
        cls,
        data: bytes,
        *,
        detail: DetailLevel | None = None,
        description: str | None = None,
    ) -> ImageBase64Content:
        """Create image content from raw bytes.

        Args:
            data: Raw image bytes
            detail: Optional detail level for processing
            description: Optional description of the image
        """
        content = base64.b64encode(data).decode()
        return cls(data=content, detail=detail, description=description)

    @classmethod
    def from_pil_image(cls, image: PIL.Image.Image) -> ImageBase64Content:
        """Create content from PIL Image."""
        with io.BytesIO() as buffer:
            image.save(buffer, format="PNG")
            return cls(data=base64.b64encode(buffer.getvalue()).decode())

data instance-attribute

data: str

Base64-encoded image data.

mime_type class-attribute instance-attribute

mime_type: str = 'image/jpeg'

MIME type of the image.

type class-attribute instance-attribute

type: Literal['image_base64'] = Field('image_base64', init=False)

Base64-encoded image.

from_bytes classmethod

from_bytes(
    data: bytes, *, detail: DetailLevel | None = None, description: str | None = None
) -> ImageBase64Content

Create image content from raw bytes.

Parameters:

Name	Type	Description	Default
data	bytes	Raw image bytes	required
detail	DetailLevel | None	Optional detail level for processing	None
description	str | None	Optional description of the image	None

Source code in src/llmling_agent/models/content.py

@classmethod
def from_bytes(
    cls,
    data: bytes,
    *,
    detail: DetailLevel | None = None,
    description: str | None = None,
) -> ImageBase64Content:
    """Create image content from raw bytes.

    Args:
        data: Raw image bytes
        detail: Optional detail level for processing
        description: Optional description of the image
    """
    content = base64.b64encode(data).decode()
    return cls(data=content, detail=detail, description=description)

from_pil_image classmethod

from_pil_image(image: Image) -> ImageBase64Content

Create content from PIL Image.

Source code in src/llmling_agent/models/content.py

@classmethod
def from_pil_image(cls, image: PIL.Image.Image) -> ImageBase64Content:
    """Create content from PIL Image."""
    with io.BytesIO() as buffer:
        image.save(buffer, format="PNG")
        return cls(data=base64.b64encode(buffer.getvalue()).decode())

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for vision models.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for vision models."""
    data_url = f"data:{self.mime_type};base64,{self.data}"
    content = {"url": data_url, "detail": self.detail or "auto"}
    return {"type": "image_url", "image_url": content}

ImageURLContent

Bases: BaseImageContent

Image from URL.

Source code in src/llmling_agent/models/content.py

class ImageURLContent(BaseImageContent):
    """Image from URL."""

    type: Literal["image_url"] = Field("image_url", init=False)
    """URL-based image."""

    url: str
    """URL to the image."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for vision models."""
        content = {"url": self.url, "detail": self.detail or "auto"}
        return {"type": "image_url", "image_url": content}

type class-attribute instance-attribute

type: Literal['image_url'] = Field('image_url', init=False)

URL-based image.

url instance-attribute

url: str

URL to the image.

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for vision models.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for vision models."""
    content = {"url": self.url, "detail": self.detail or "auto"}
    return {"type": "image_url", "image_url": content}

MessageNode

Bases: MessageEmitter[TDeps, TResult]

Base class for all message processing nodes.

Source code in src/llmling_agent/messaging/messagenode.py

class MessageNode[TDeps, TResult](MessageEmitter[TDeps, TResult]):
    """Base class for all message processing nodes."""

    tool_used = Signal(ToolCallInfo)
    """Signal emitted when node uses a tool."""

    async def pre_run(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str] | ChatMessage,
    ) -> tuple[ChatMessage[Any], list[Content | str]]:
        """Hook to prepare a MessgeNode run call.

        Args:
            *prompt: The prompt(s) to prepare.

        Returns:
            A tuple of:
                - Either incoming message, or a constructed incoming message based
                  on the prompt(s).
                - A list of prompts to be sent to the model.
        """
        if len(prompt) == 1 and isinstance(prompt[0], ChatMessage):
            user_msg = prompt[0]
            prompts = await convert_prompts([user_msg.content])
            # Update received message's chain to show it came through its source
            user_msg = user_msg.forwarded(prompt[0])
            # clear cost info to avoid double-counting
            user_msg = replace(user_msg, role="user", cost_info=None)
            final_prompt = "\n\n".join(str(p) for p in prompts)
        else:
            prompts = await convert_prompts(prompt)
            final_prompt = "\n\n".join(str(p) for p in prompts)
            # use format_prompts?
            user_msg = ChatMessage[str](
                content=final_prompt,
                role="user",
                conversation_id=str(uuid4()),
            )
        self.message_received.emit(user_msg)
        self.context.current_prompt = final_prompt
        return user_msg, prompts

    # async def post_run(
    #     self,
    #     message: ChatMessage[TResult],
    #     previous_message: ChatMessage[Any] | None,
    #     wait_for_connections: bool | None = None,
    # ) -> ChatMessage[Any]:
    #     # For chain processing, update the response's chain
    #     if previous_message:
    #         message = message.forwarded(previous_message)
    #         conversation_id = previous_message.conversation_id
    #     else:
    #         conversation_id = str(uuid4())
    #     # Set conversation_id on response message
    #     message = replace(message, conversation_id=conversation_id)
    #     self.message_sent.emit(message)
    #     await self.connections.route_message(message, wait=wait_for_connections)
    #     return message

    async def run(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str] | ChatMessage,
        wait_for_connections: bool | None = None,
        store_history: bool = True,
        **kwargs: Any,
    ) -> ChatMessage[TResult]:
        """Execute node with prompts and handle message routing.

        Args:
            prompt: Input prompts
            wait_for_connections: Whether to wait for forwarded messages
            store_history: Whether to store in conversation history
            **kwargs: Additional arguments for _run
        """
        from llmling_agent import Agent, StructuredAgent

        user_msg, prompts = await self.pre_run(*prompt)
        message = await self._run(
            *prompts,
            store_history=store_history,
            conversation_id=user_msg.conversation_id,
            **kwargs,
        )

        # For chain processing, update the response's chain
        if len(prompt) == 1 and isinstance(prompt[0], ChatMessage):
            message = message.forwarded(prompt[0])

        if store_history and isinstance(self, Agent | StructuredAgent):
            self.conversation.add_chat_messages([user_msg, message])
        self.message_sent.emit(message)
        await self.connections.route_message(message, wait=wait_for_connections)
        return message

    @abstractmethod
    def run_iter(
        self,
        *prompts: Any,
        **kwargs: Any,
    ) -> AsyncIterator[ChatMessage[Any]]:
        """Yield messages during execution."""

tool_used class-attribute instance-attribute

tool_used = Signal(ToolCallInfo)

Signal emitted when node uses a tool.

pre_run async

pre_run(
    *prompt: AnyPromptType | Image | PathLike[str] | ChatMessage,
) -> tuple[ChatMessage[Any], list[Content | str]]

Hook to prepare a MessgeNode run call.

Parameters:

Name	Type	Description	Default
*prompt	AnyPromptType | Image | PathLike[str] | ChatMessage	The prompt(s) to prepare.	()

Returns:

Type	Description
tuple[ChatMessage[Any], list[Content | str]]	A tuple of: - Either incoming message, or a constructed incoming message based on the prompt(s). - A list of prompts to be sent to the model.

Source code in src/llmling_agent/messaging/messagenode.py

async def pre_run(
    self,
    *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str] | ChatMessage,
) -> tuple[ChatMessage[Any], list[Content | str]]:
    """Hook to prepare a MessgeNode run call.

    Args:
        *prompt: The prompt(s) to prepare.

    Returns:
        A tuple of:
            - Either incoming message, or a constructed incoming message based
              on the prompt(s).
            - A list of prompts to be sent to the model.
    """
    if len(prompt) == 1 and isinstance(prompt[0], ChatMessage):
        user_msg = prompt[0]
        prompts = await convert_prompts([user_msg.content])
        # Update received message's chain to show it came through its source
        user_msg = user_msg.forwarded(prompt[0])
        # clear cost info to avoid double-counting
        user_msg = replace(user_msg, role="user", cost_info=None)
        final_prompt = "\n\n".join(str(p) for p in prompts)
    else:
        prompts = await convert_prompts(prompt)
        final_prompt = "\n\n".join(str(p) for p in prompts)
        # use format_prompts?
        user_msg = ChatMessage[str](
            content=final_prompt,
            role="user",
            conversation_id=str(uuid4()),
        )
    self.message_received.emit(user_msg)
    self.context.current_prompt = final_prompt
    return user_msg, prompts

run async

run(
    *prompt: AnyPromptType | Image | PathLike[str] | ChatMessage,
    wait_for_connections: bool | None = None,
    store_history: bool = True,
    **kwargs: Any,
) -> ChatMessage[TResult]

Execute node with prompts and handle message routing.

Parameters:

Name	Type	Description	Default
prompt	AnyPromptType | Image | PathLike[str] | ChatMessage	Input prompts	()
wait_for_connections	bool | None	Whether to wait for forwarded messages	None
store_history	bool	Whether to store in conversation history	True
**kwargs	Any	Additional arguments for _run	{}

Source code in src/llmling_agent/messaging/messagenode.py

async def run(
    self,
    *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str] | ChatMessage,
    wait_for_connections: bool | None = None,
    store_history: bool = True,
    **kwargs: Any,
) -> ChatMessage[TResult]:
    """Execute node with prompts and handle message routing.

    Args:
        prompt: Input prompts
        wait_for_connections: Whether to wait for forwarded messages
        store_history: Whether to store in conversation history
        **kwargs: Additional arguments for _run
    """
    from llmling_agent import Agent, StructuredAgent

    user_msg, prompts = await self.pre_run(*prompt)
    message = await self._run(
        *prompts,
        store_history=store_history,
        conversation_id=user_msg.conversation_id,
        **kwargs,
    )

    # For chain processing, update the response's chain
    if len(prompt) == 1 and isinstance(prompt[0], ChatMessage):
        message = message.forwarded(prompt[0])

    if store_history and isinstance(self, Agent | StructuredAgent):
        self.conversation.add_chat_messages([user_msg, message])
    self.message_sent.emit(message)
    await self.connections.route_message(message, wait=wait_for_connections)
    return message

run_iter abstractmethod

run_iter(*prompts: Any, **kwargs: Any) -> AsyncIterator[ChatMessage[Any]]

Yield messages during execution.

Source code in src/llmling_agent/messaging/messagenode.py

@abstractmethod
def run_iter(
    self,
    *prompts: Any,
    **kwargs: Any,
) -> AsyncIterator[ChatMessage[Any]]:
    """Yield messages during execution."""

PDFBase64Content

Bases: BasePDFContent

PDF from base64 data.

Source code in src/llmling_agent/models/content.py

class PDFBase64Content(BasePDFContent):
    """PDF from base64 data."""

    type: Literal["pdf_base64"] = Field("pdf_base64", init=False)
    """Base64-data based PDF."""

    data: str
    """Base64-encoded PDF data."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for PDF handling."""
        data_url = f"data:application/pdf;base64,{self.data}"
        content = {"url": data_url, "detail": self.detail or "auto"}
        return {"type": "file", "file": content}

    @classmethod
    def from_bytes(
        cls,
        data: bytes,
        *,
        detail: DetailLevel | None = None,
        description: str | None = None,
    ) -> Self:
        """Create PDF content from raw bytes."""
        content = base64.b64encode(data).decode()
        return cls(data=content, detail=detail, description=description)

data instance-attribute

data: str

Base64-encoded PDF data.

type class-attribute instance-attribute

type: Literal['pdf_base64'] = Field('pdf_base64', init=False)

Base64-data based PDF.

from_bytes classmethod

from_bytes(
    data: bytes, *, detail: DetailLevel | None = None, description: str | None = None
) -> Self

Create PDF content from raw bytes.

Source code in src/llmling_agent/models/content.py

@classmethod
def from_bytes(
    cls,
    data: bytes,
    *,
    detail: DetailLevel | None = None,
    description: str | None = None,
) -> Self:
    """Create PDF content from raw bytes."""
    content = base64.b64encode(data).decode()
    return cls(data=content, detail=detail, description=description)

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for PDF handling.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for PDF handling."""
    data_url = f"data:application/pdf;base64,{self.data}"
    content = {"url": data_url, "detail": self.detail or "auto"}
    return {"type": "file", "file": content}

PDFURLContent

Bases: BasePDFContent

PDF from URL.

Source code in src/llmling_agent/models/content.py

class PDFURLContent(BasePDFContent):
    """PDF from URL."""

    type: Literal["pdf_url"] = Field("pdf_url", init=False)
    """URL-based PDF."""

    url: str
    """URL to the PDF document."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for PDF handling."""
        content = {"url": self.url, "detail": self.detail or "auto"}
        return {"type": "file", "file": content}

type class-attribute instance-attribute

type: Literal['pdf_url'] = Field('pdf_url', init=False)

URL-based PDF.

url instance-attribute

url: str

URL to the PDF document.

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for PDF handling.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for PDF handling."""
    content = {"url": self.url, "detail": self.detail or "auto"}
    return {"type": "file", "file": content}

StructuredAgent

Bases: MessageNode[TDeps, TResult]

Wrapper for Agent that enforces a specific result type.

This wrapper ensures the agent always returns results of the specified type. The type can be provided as: - A Python type for validation - A response definition name from the manifest - A complete response definition instance

Source code in src/llmling_agent/agent/structured.py

class StructuredAgent[TDeps, TResult](MessageNode[TDeps, TResult]):
    """Wrapper for Agent that enforces a specific result type.

    This wrapper ensures the agent always returns results of the specified type.
    The type can be provided as:
    - A Python type for validation
    - A response definition name from the manifest
    - A complete response definition instance
    """

    def __init__(
        self,
        agent: Agent[TDeps] | StructuredAgent[TDeps, TResult] | Callable[..., TResult],
        result_type: type[TResult] | str | ResponseDefinition,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ):
        """Initialize structured agent wrapper.

        Args:
            agent: Base agent to wrap
            result_type: Expected result type:
                - BaseModel / dataclasses
                - Name of response definition in manifest
                - Complete response definition instance
            tool_name: Optional override for tool name
            tool_description: Optional override for tool description

        Raises:
            ValueError: If named response type not found in manifest
        """
        from llmling_agent.agent.agent import Agent

        logger.debug("StructuredAgent.run result_type = %s", result_type)
        match agent:
            case StructuredAgent():
                self._agent: Agent[TDeps] = agent._agent
            case Callable():
                self._agent = Agent[TDeps](provider=agent, name=agent.__name__)
            case Agent():
                self._agent = agent
            case _:
                msg = "Invalid agent type"
                raise ValueError(msg)

        super().__init__(name=self._agent.name)

        self._result_type = to_type(result_type)
        agent.set_result_type(result_type)

        match result_type:
            case type() | str():
                # For types and named definitions, use overrides if provided
                self._agent.set_result_type(
                    result_type,
                    tool_name=tool_name,
                    tool_description=tool_description,
                )
            case BaseResponseDefinition():
                # For response definitions, use as-is
                # (overrides don't apply to complete definitions)
                self._agent.set_result_type(result_type)

    async def __aenter__(self) -> Self:
        """Enter async context and set up MCP servers.

        Called when agent enters its async context. Sets up any configured
        MCP servers and their tools.
        """
        await self._agent.__aenter__()
        return self

    async def __aexit__(
        self,
        exc_type: type[BaseException] | None,
        exc_val: BaseException | None,
        exc_tb: TracebackType | None,
    ):
        """Exit async context."""
        await self._agent.__aexit__(exc_type, exc_val, exc_tb)

    def __and__(
        self, other: AnyAgent[Any, Any] | Team[Any] | ProcessorCallback[TResult]
    ) -> Team[TDeps]:
        return self._agent.__and__(other)

    def __or__(self, other: Agent | ProcessorCallback | BaseTeam) -> TeamRun:
        return self._agent.__or__(other)

    async def _run(
        self,
        *prompt: AnyPromptType | TResult,
        result_type: type[TResult] | None = None,
        model: ModelType = None,
        tool_choice: str | list[str] | None = None,
        store_history: bool = True,
        message_id: str | None = None,
        conversation_id: str | None = None,
        wait_for_connections: bool | None = None,
    ) -> ChatMessage[TResult]:
        """Run with fixed result type.

        Args:
            prompt: Any prompt-compatible object or structured objects of type TResult
            result_type: Expected result type:
                - BaseModel / dataclasses
                - Name of response definition in manifest
                - Complete response definition instance
            model: Optional model override
            tool_choice: Filter available tools by name
            store_history: Whether the message exchange should be added to the
                           context window
            message_id: Optional message id for the returned message.
                        Automatically generated if not provided.
            conversation_id: Optional conversation id for the returned message.
            wait_for_connections: Whether to wait for all connections to complete
        """
        typ = result_type or self._result_type
        return await self._agent._run(
            *prompt,
            result_type=typ,
            model=model,
            store_history=store_history,
            tool_choice=tool_choice,
            message_id=message_id,
            conversation_id=conversation_id,
            wait_for_connections=wait_for_connections,
        )

    async def validate_against(
        self,
        prompt: str,
        criteria: type[TResult],
        **kwargs: Any,
    ) -> bool:
        """Check if agent's response satisfies stricter criteria."""
        result = await self.run(prompt, **kwargs)
        try:
            criteria.model_validate(result.content.model_dump())  # type: ignore
        except ValidationError:
            return False
        else:
            return True

    def __repr__(self) -> str:
        type_name = getattr(self._result_type, "__name__", str(self._result_type))
        return f"StructuredAgent({self._agent!r}, result_type={type_name})"

    def __prompt__(self) -> str:
        type_name = getattr(self._result_type, "__name__", str(self._result_type))
        base_info = self._agent.__prompt__()
        return f"{base_info}\nStructured output type: {type_name}"

    def __getattr__(self, name: str) -> Any:
        return getattr(self._agent, name)

    @property
    def context(self) -> AgentContext[TDeps]:
        return self._agent.context

    @context.setter
    def context(self, value: Any):
        self._agent.context = value

    @property
    def name(self) -> str:
        return self._agent.name

    @name.setter
    def name(self, value: str):
        self._agent.name = value

    @property
    def tools(self) -> ToolManager:
        return self._agent.tools

    @property
    def conversation(self) -> ConversationManager:
        return self._agent.conversation

    @overload
    def to_structured(
        self,
        result_type: None,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ) -> Agent[TDeps]: ...

    @overload
    def to_structured[TNewResult](
        self,
        result_type: type[TNewResult] | str | ResponseDefinition,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ) -> StructuredAgent[TDeps, TNewResult]: ...

    def to_structured[TNewResult](
        self,
        result_type: type[TNewResult] | str | ResponseDefinition | None,
        *,
        tool_name: str | None = None,
        tool_description: str | None = None,
    ) -> Agent[TDeps] | StructuredAgent[TDeps, TNewResult]:
        if result_type is None:
            return self._agent

        return StructuredAgent(
            self._agent,
            result_type=result_type,
            tool_name=tool_name,
            tool_description=tool_description,
        )

    @property
    def stats(self) -> MessageStats:
        return self._agent.stats

    async def run_iter(
        self,
        *prompt_groups: Sequence[AnyPromptType | PIL.Image.Image | os.PathLike[str]],
        **kwargs: Any,
    ) -> AsyncIterator[ChatMessage[Any]]:
        """Forward run_iter to wrapped agent."""
        async for message in self._agent.run_iter(*prompt_groups, **kwargs):
            yield message

    async def run_job(
        self,
        job: Job[TDeps, TResult],
        *,
        store_history: bool = True,
        include_agent_tools: bool = True,
    ) -> ChatMessage[TResult]:
        """Execute a pre-defined job ensuring type compatibility.

        Args:
            job: Job configuration to execute
            store_history: Whether to add job execution to conversation history
            include_agent_tools: Whether to include agent's tools alongside job tools

        Returns:
            Task execution result

        Raises:
            JobError: If job execution fails or types don't match
            ValueError: If job configuration is invalid
        """
        from llmling_agent.tasks import JobError

        # Validate dependency requirement
        if job.required_dependency is not None:  # noqa: SIM102
            if not isinstance(self.context.data, job.required_dependency):
                msg = (
                    f"Agent dependencies ({type(self.context.data)}) "
                    f"don't match job requirement ({job.required_dependency})"
                )
                raise JobError(msg)

        # Validate return type requirement
        if job.required_return_type != self._result_type:
            msg = (
                f"Agent result type ({self._result_type}) "
                f"doesn't match job requirement ({job.required_return_type})"
            )
            raise JobError(msg)

        # Load task knowledge if provided
        if job.knowledge:
            # Add knowledge sources to context
            resources: list[Resource | str] = list(job.knowledge.paths) + list(
                job.knowledge.resources
            )
            for source in resources:
                await self.conversation.load_context_source(source)
            for prompt in job.knowledge.prompts:
                await self.conversation.load_context_source(prompt)

        try:
            # Register task tools temporarily
            tools = job.get_tools()

            # Use temporary tools
            with self._agent.tools.temporary_tools(
                tools, exclusive=not include_agent_tools
            ):
                # Execute job using StructuredAgent's run to maintain type safety
                return await self.run(await job.get_prompt(), store_history=store_history)

        except Exception as e:
            msg = f"Task execution failed: {e}"
            logger.exception(msg)
            raise JobError(msg) from e

    @classmethod
    def from_callback(
        cls,
        callback: ProcessorCallback[TResult],
        *,
        name: str | None = None,
        **kwargs: Any,
    ) -> StructuredAgent[None, TResult]:
        """Create a structured agent from a processing callback.

        Args:
            callback: Function to process messages. Can be:
                - sync or async
                - with or without context
                - with explicit return type
            name: Optional name for the agent
            **kwargs: Additional arguments for agent

        Example:
            ```python
            class AnalysisResult(BaseModel):
                sentiment: float
                topics: list[str]

            def analyze(msg: str) -> AnalysisResult:
                return AnalysisResult(sentiment=0.8, topics=["tech"])

            analyzer = StructuredAgent.from_callback(analyze)
            ```
        """
        from llmling_agent.agent.agent import Agent
        from llmling_agent_providers.callback import CallbackProvider

        name = name or callback.__name__ or "processor"
        provider = CallbackProvider(callback, name=name)
        agent = Agent[None](provider=provider, name=name, **kwargs)
        # Get return type from signature for validation
        hints = get_type_hints(callback)
        return_type = hints.get("return")

        # If async, unwrap from Awaitable
        if (
            return_type
            and hasattr(return_type, "__origin__")
            and return_type.__origin__ is Awaitable
        ):
            return_type = return_type.__args__[0]
        return StructuredAgent[None, TResult](agent, return_type or str)  # type: ignore

    def is_busy(self) -> bool:
        """Check if agent is currently processing tasks."""
        return bool(self._pending_tasks or self._background_task)

    def run_sync(self, *args, **kwargs):
        """Run agent synchronously."""
        return self._agent.run_sync(*args, result_type=self._result_type, **kwargs)

__aenter__ async

__aenter__() -> Self

Enter async context and set up MCP servers.

Called when agent enters its async context. Sets up any configured MCP servers and their tools.

Source code in src/llmling_agent/agent/structured.py

async def __aenter__(self) -> Self:
    """Enter async context and set up MCP servers.

    Called when agent enters its async context. Sets up any configured
    MCP servers and their tools.
    """
    await self._agent.__aenter__()
    return self

__aexit__ async

__aexit__(
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
)

Exit async context.

Source code in src/llmling_agent/agent/structured.py

async def __aexit__(
    self,
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
):
    """Exit async context."""
    await self._agent.__aexit__(exc_type, exc_val, exc_tb)

__init__

__init__(
    agent: Agent[TDeps] | StructuredAgent[TDeps, TResult] | Callable[..., TResult],
    result_type: type[TResult] | str | ResponseDefinition,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
)

Initialize structured agent wrapper.

Parameters:

Name	Type	Description	Default
agent	Agent[TDeps] | StructuredAgent[TDeps, TResult] | Callable[..., TResult]	Base agent to wrap	required
result_type	type[TResult] | str | ResponseDefinition	Expected result type: - BaseModel / dataclasses - Name of response definition in manifest - Complete response definition instance	required
tool_name	str | None	Optional override for tool name	None
tool_description	str | None	Optional override for tool description	None

Raises:

Type	Description
ValueError	If named response type not found in manifest

Source code in src/llmling_agent/agent/structured.py

def __init__(
    self,
    agent: Agent[TDeps] | StructuredAgent[TDeps, TResult] | Callable[..., TResult],
    result_type: type[TResult] | str | ResponseDefinition,
    *,
    tool_name: str | None = None,
    tool_description: str | None = None,
):
    """Initialize structured agent wrapper.

    Args:
        agent: Base agent to wrap
        result_type: Expected result type:
            - BaseModel / dataclasses
            - Name of response definition in manifest
            - Complete response definition instance
        tool_name: Optional override for tool name
        tool_description: Optional override for tool description

    Raises:
        ValueError: If named response type not found in manifest
    """
    from llmling_agent.agent.agent import Agent

    logger.debug("StructuredAgent.run result_type = %s", result_type)
    match agent:
        case StructuredAgent():
            self._agent: Agent[TDeps] = agent._agent
        case Callable():
            self._agent = Agent[TDeps](provider=agent, name=agent.__name__)
        case Agent():
            self._agent = agent
        case _:
            msg = "Invalid agent type"
            raise ValueError(msg)

    super().__init__(name=self._agent.name)

    self._result_type = to_type(result_type)
    agent.set_result_type(result_type)

    match result_type:
        case type() | str():
            # For types and named definitions, use overrides if provided
            self._agent.set_result_type(
                result_type,
                tool_name=tool_name,
                tool_description=tool_description,
            )
        case BaseResponseDefinition():
            # For response definitions, use as-is
            # (overrides don't apply to complete definitions)
            self._agent.set_result_type(result_type)

from_callback classmethod

from_callback(
    callback: ProcessorCallback[TResult], *, name: str | None = None, **kwargs: Any
) -> StructuredAgent[None, TResult]

Create a structured agent from a processing callback.

Parameters:

Name	Type	Description	Default
callback	ProcessorCallback[TResult]	Function to process messages. Can be: - sync or async - with or without context - with explicit return type	required
name	str | None	Optional name for the agent	None
**kwargs	Any	Additional arguments for agent	{}

Example

class AnalysisResult(BaseModel):
    sentiment: float
    topics: list[str]

def analyze(msg: str) -> AnalysisResult:
    return AnalysisResult(sentiment=0.8, topics=["tech"])

analyzer = StructuredAgent.from_callback(analyze)

Source code in src/llmling_agent/agent/structured.py

@classmethod
def from_callback(
    cls,
    callback: ProcessorCallback[TResult],
    *,
    name: str | None = None,
    **kwargs: Any,
) -> StructuredAgent[None, TResult]:
    """Create a structured agent from a processing callback.

    Args:
        callback: Function to process messages. Can be:
            - sync or async
            - with or without context
            - with explicit return type
        name: Optional name for the agent
        **kwargs: Additional arguments for agent

    Example:
        ```python
        class AnalysisResult(BaseModel):
            sentiment: float
            topics: list[str]

        def analyze(msg: str) -> AnalysisResult:
            return AnalysisResult(sentiment=0.8, topics=["tech"])

        analyzer = StructuredAgent.from_callback(analyze)
        ```
    """
    from llmling_agent.agent.agent import Agent
    from llmling_agent_providers.callback import CallbackProvider

    name = name or callback.__name__ or "processor"
    provider = CallbackProvider(callback, name=name)
    agent = Agent[None](provider=provider, name=name, **kwargs)
    # Get return type from signature for validation
    hints = get_type_hints(callback)
    return_type = hints.get("return")

    # If async, unwrap from Awaitable
    if (
        return_type
        and hasattr(return_type, "__origin__")
        and return_type.__origin__ is Awaitable
    ):
        return_type = return_type.__args__[0]
    return StructuredAgent[None, TResult](agent, return_type or str)  # type: ignore

is_busy

is_busy() -> bool

Check if agent is currently processing tasks.

Source code in src/llmling_agent/agent/structured.py

def is_busy(self) -> bool:
    """Check if agent is currently processing tasks."""
    return bool(self._pending_tasks or self._background_task)

run_iter async

run_iter(
    *prompt_groups: Sequence[AnyPromptType | Image | PathLike[str]], **kwargs: Any
) -> AsyncIterator[ChatMessage[Any]]

Forward run_iter to wrapped agent.

Source code in src/llmling_agent/agent/structured.py

async def run_iter(
    self,
    *prompt_groups: Sequence[AnyPromptType | PIL.Image.Image | os.PathLike[str]],
    **kwargs: Any,
) -> AsyncIterator[ChatMessage[Any]]:
    """Forward run_iter to wrapped agent."""
    async for message in self._agent.run_iter(*prompt_groups, **kwargs):
        yield message

run_job async

run_job(
    job: Job[TDeps, TResult],
    *,
    store_history: bool = True,
    include_agent_tools: bool = True,
) -> ChatMessage[TResult]

Execute a pre-defined job ensuring type compatibility.

Parameters:

Name	Type	Description	Default
job	Job[TDeps, TResult]	Job configuration to execute	required
store_history	bool	Whether to add job execution to conversation history	True
include_agent_tools	bool	Whether to include agent's tools alongside job tools	True

Returns:

Type	Description
ChatMessage[TResult]	Task execution result

Raises:

Type	Description
JobError	If job execution fails or types don't match
ValueError	If job configuration is invalid

Source code in src/llmling_agent/agent/structured.py

async def run_job(
    self,
    job: Job[TDeps, TResult],
    *,
    store_history: bool = True,
    include_agent_tools: bool = True,
) -> ChatMessage[TResult]:
    """Execute a pre-defined job ensuring type compatibility.

    Args:
        job: Job configuration to execute
        store_history: Whether to add job execution to conversation history
        include_agent_tools: Whether to include agent's tools alongside job tools

    Returns:
        Task execution result

    Raises:
        JobError: If job execution fails or types don't match
        ValueError: If job configuration is invalid
    """
    from llmling_agent.tasks import JobError

    # Validate dependency requirement
    if job.required_dependency is not None:  # noqa: SIM102
        if not isinstance(self.context.data, job.required_dependency):
            msg = (
                f"Agent dependencies ({type(self.context.data)}) "
                f"don't match job requirement ({job.required_dependency})"
            )
            raise JobError(msg)

    # Validate return type requirement
    if job.required_return_type != self._result_type:
        msg = (
            f"Agent result type ({self._result_type}) "
            f"doesn't match job requirement ({job.required_return_type})"
        )
        raise JobError(msg)

    # Load task knowledge if provided
    if job.knowledge:
        # Add knowledge sources to context
        resources: list[Resource | str] = list(job.knowledge.paths) + list(
            job.knowledge.resources
        )
        for source in resources:
            await self.conversation.load_context_source(source)
        for prompt in job.knowledge.prompts:
            await self.conversation.load_context_source(prompt)

    try:
        # Register task tools temporarily
        tools = job.get_tools()

        # Use temporary tools
        with self._agent.tools.temporary_tools(
            tools, exclusive=not include_agent_tools
        ):
            # Execute job using StructuredAgent's run to maintain type safety
            return await self.run(await job.get_prompt(), store_history=store_history)

    except Exception as e:
        msg = f"Task execution failed: {e}"
        logger.exception(msg)
        raise JobError(msg) from e

run_sync

run_sync(*args, **kwargs)

Run agent synchronously.

Source code in src/llmling_agent/agent/structured.py

def run_sync(self, *args, **kwargs):
    """Run agent synchronously."""
    return self._agent.run_sync(*args, result_type=self._result_type, **kwargs)

validate_against async

validate_against(prompt: str, criteria: type[TResult], **kwargs: Any) -> bool

Check if agent's response satisfies stricter criteria.

Source code in src/llmling_agent/agent/structured.py

async def validate_against(
    self,
    prompt: str,
    criteria: type[TResult],
    **kwargs: Any,
) -> bool:
    """Check if agent's response satisfies stricter criteria."""
    result = await self.run(prompt, **kwargs)
    try:
        criteria.model_validate(result.content.model_dump())  # type: ignore
    except ValidationError:
        return False
    else:
        return True

Team

Bases: BaseTeam[TDeps, Any]

Group of agents that can execute together.

Source code in src/llmling_agent/delegation/team.py

class Team[TDeps](BaseTeam[TDeps, Any]):
    """Group of agents that can execute together."""

    async def execute(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        **kwargs: Any,
    ) -> TeamResponse:
        """Run all agents in parallel with monitoring."""
        from llmling_agent.talk.talk import Talk

        self._team_talk.clear()

        start_time = get_now()
        responses: list[AgentResponse[Any]] = []
        errors: dict[str, Exception] = {}
        final_prompt = list(prompts)
        if self.shared_prompt:
            final_prompt.insert(0, self.shared_prompt)
        combined_prompt = "\n".join([await to_prompt(p) for p in final_prompt])
        all_nodes = list(await self.pick_agents(combined_prompt))
        # Create Talk connections for monitoring this execution
        execution_talks: list[Talk[Any]] = []
        for node in all_nodes:
            talk = Talk[Any](
                node,
                [],  # No actual forwarding, just for tracking
                connection_type="run",
                queued=True,
                queue_strategy="latest",
            )
            execution_talks.append(talk)
            self._team_talk.append(talk)  # Add to base class's TeamTalk

        async def _run(node: MessageNode[TDeps, Any]):
            try:
                start = perf_counter()
                message = await node.run(*final_prompt, **kwargs)
                timing = perf_counter() - start
                r = AgentResponse(agent_name=node.name, message=message, timing=timing)
                responses.append(r)

                # Update talk stats for this agent
                talk = next(t for t in execution_talks if t.source == node)
                talk._stats.messages.append(message)

            except Exception as e:  # noqa: BLE001
                errors[node.name] = e

        # Run all agents in parallel
        await asyncio.gather(*[_run(node) for node in all_nodes])

        return TeamResponse(responses=responses, start_time=start_time, errors=errors)

    def __prompt__(self) -> str:
        """Format team info for prompts."""
        members = ", ".join(a.name for a in self.agents)
        desc = f" - {self.description}" if self.description else ""
        return f"Parallel Team '{self.name}'{desc}\nMembers: {members}"

    async def run_iter(
        self,
        *prompts: AnyPromptType,
        **kwargs: Any,
    ) -> AsyncIterator[ChatMessage[Any]]:
        """Yield messages as they arrive from parallel execution."""
        queue: asyncio.Queue[ChatMessage[Any] | None] = asyncio.Queue()
        failures: dict[str, Exception] = {}

        async def _run(node: MessageNode[TDeps, Any]):
            try:
                message = await node.run(*prompts, **kwargs)
                await queue.put(message)
            except Exception as e:
                logger.exception("Error executing node %s", node.name)
                failures[node.name] = e
                # Put None to maintain queue count
                await queue.put(None)

        # Get nodes to run
        combined_prompt = "\n".join([await to_prompt(p) for p in prompts])
        all_nodes = list(await self.pick_agents(combined_prompt))

        # Start all agents
        tasks = [asyncio.create_task(_run(n), name=f"run_{n.name}") for n in all_nodes]

        try:
            # Yield messages as they arrive
            for _ in all_nodes:
                if msg := await queue.get():
                    yield msg

            # If any failures occurred, raise error with details
            if failures:
                error_details = "\n".join(
                    f"- {name}: {error}" for name, error in failures.items()
                )
                error_msg = f"Some nodes failed to execute:\n{error_details}"
                raise RuntimeError(error_msg)

        finally:
            # Clean up any remaining tasks
            for task in tasks:
                if not task.done():
                    task.cancel()

    async def _run(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        wait_for_connections: bool | None = None,
        message_id: str | None = None,
        conversation_id: str | None = None,
        **kwargs: Any,
    ) -> ChatMessage[list[Any]]:
        """Run all agents in parallel and return combined message."""
        result: TeamResponse = await self.execute(*prompts, **kwargs)
        message_id = message_id or str(uuid4())
        return ChatMessage(
            content=[r.message.content for r in result if r.message],
            role="assistant",
            name=self.name,
            message_id=message_id,
            conversation_id=conversation_id,
            metadata={
                "agent_names": [r.agent_name for r in result],
                "errors": {name: str(error) for name, error in result.errors.items()},
                "start_time": result.start_time.isoformat(),
            },
        )

    async def run_job[TJobResult](
        self,
        job: Job[TDeps, TJobResult],
        *,
        store_history: bool = True,
        include_agent_tools: bool = True,
    ) -> list[AgentResponse[TJobResult]]:
        """Execute a job across all team members in parallel.

        Args:
            job: Job configuration to execute
            store_history: Whether to add job execution to conversation history
            include_agent_tools: Whether to include agent's tools alongside job tools

        Returns:
            List of responses from all agents

        Raises:
            JobError: If job execution fails for any agent
            ValueError: If job configuration is invalid
        """
        from llmling_agent.agent import Agent, StructuredAgent
        from llmling_agent.tasks import JobError

        responses: list[AgentResponse[TJobResult]] = []
        errors: dict[str, Exception] = {}
        start_time = get_now()

        # Validate dependencies for all agents
        if job.required_dependency is not None:
            invalid_agents = [
                agent.name
                for agent in self.iter_agents()
                if not isinstance(agent.context.data, job.required_dependency)
            ]
            if invalid_agents:
                msg = (
                    f"Agents {', '.join(invalid_agents)} don't have required "
                    f"dependency type: {job.required_dependency}"
                )
                raise JobError(msg)

        try:
            # Load knowledge for all agents if provided
            if job.knowledge:
                # TODO: resources
                tools = [t.name for t in job.get_tools()]
                await self.distribute(content="", tools=tools)

            prompt = await job.get_prompt()

            async def _run(agent: MessageNode[TDeps, TJobResult]):
                assert isinstance(agent, Agent | StructuredAgent)
                try:
                    with agent.tools.temporary_tools(
                        job.get_tools(), exclusive=not include_agent_tools
                    ):
                        start = perf_counter()
                        resp = AgentResponse(
                            agent_name=agent.name,
                            message=await agent.run(prompt, store_history=store_history),  # pyright: ignore
                            timing=perf_counter() - start,
                        )
                        responses.append(resp)
                except Exception as e:  # noqa: BLE001
                    errors[agent.name] = e

            # Run job in parallel on all agents
            await asyncio.gather(*[_run(node) for node in self.agents])

            return TeamResponse(responses=responses, start_time=start_time, errors=errors)

        except Exception as e:
            msg = "Job execution failed"
            logger.exception(msg)
            raise JobError(msg) from e

__prompt__

__prompt__() -> str

Format team info for prompts.

Source code in src/llmling_agent/delegation/team.py

def __prompt__(self) -> str:
    """Format team info for prompts."""
    members = ", ".join(a.name for a in self.agents)
    desc = f" - {self.description}" if self.description else ""
    return f"Parallel Team '{self.name}'{desc}\nMembers: {members}"

execute async

execute(
    *prompts: AnyPromptType | Image | PathLike[str] | None, **kwargs: Any
) -> TeamResponse

Run all agents in parallel with monitoring.

Source code in src/llmling_agent/delegation/team.py

async def execute(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
    **kwargs: Any,
) -> TeamResponse:
    """Run all agents in parallel with monitoring."""
    from llmling_agent.talk.talk import Talk

    self._team_talk.clear()

    start_time = get_now()
    responses: list[AgentResponse[Any]] = []
    errors: dict[str, Exception] = {}
    final_prompt = list(prompts)
    if self.shared_prompt:
        final_prompt.insert(0, self.shared_prompt)
    combined_prompt = "\n".join([await to_prompt(p) for p in final_prompt])
    all_nodes = list(await self.pick_agents(combined_prompt))
    # Create Talk connections for monitoring this execution
    execution_talks: list[Talk[Any]] = []
    for node in all_nodes:
        talk = Talk[Any](
            node,
            [],  # No actual forwarding, just for tracking
            connection_type="run",
            queued=True,
            queue_strategy="latest",
        )
        execution_talks.append(talk)
        self._team_talk.append(talk)  # Add to base class's TeamTalk

    async def _run(node: MessageNode[TDeps, Any]):
        try:
            start = perf_counter()
            message = await node.run(*final_prompt, **kwargs)
            timing = perf_counter() - start
            r = AgentResponse(agent_name=node.name, message=message, timing=timing)
            responses.append(r)

            # Update talk stats for this agent
            talk = next(t for t in execution_talks if t.source == node)
            talk._stats.messages.append(message)

        except Exception as e:  # noqa: BLE001
            errors[node.name] = e

    # Run all agents in parallel
    await asyncio.gather(*[_run(node) for node in all_nodes])

    return TeamResponse(responses=responses, start_time=start_time, errors=errors)

run_iter async

run_iter(*prompts: AnyPromptType, **kwargs: Any) -> AsyncIterator[ChatMessage[Any]]

Yield messages as they arrive from parallel execution.

Source code in src/llmling_agent/delegation/team.py

async def run_iter(
    self,
    *prompts: AnyPromptType,
    **kwargs: Any,
) -> AsyncIterator[ChatMessage[Any]]:
    """Yield messages as they arrive from parallel execution."""
    queue: asyncio.Queue[ChatMessage[Any] | None] = asyncio.Queue()
    failures: dict[str, Exception] = {}

    async def _run(node: MessageNode[TDeps, Any]):
        try:
            message = await node.run(*prompts, **kwargs)
            await queue.put(message)
        except Exception as e:
            logger.exception("Error executing node %s", node.name)
            failures[node.name] = e
            # Put None to maintain queue count
            await queue.put(None)

    # Get nodes to run
    combined_prompt = "\n".join([await to_prompt(p) for p in prompts])
    all_nodes = list(await self.pick_agents(combined_prompt))

    # Start all agents
    tasks = [asyncio.create_task(_run(n), name=f"run_{n.name}") for n in all_nodes]

    try:
        # Yield messages as they arrive
        for _ in all_nodes:
            if msg := await queue.get():
                yield msg

        # If any failures occurred, raise error with details
        if failures:
            error_details = "\n".join(
                f"- {name}: {error}" for name, error in failures.items()
            )
            error_msg = f"Some nodes failed to execute:\n{error_details}"
            raise RuntimeError(error_msg)

    finally:
        # Clean up any remaining tasks
        for task in tasks:
            if not task.done():
                task.cancel()

run_job async

run_job(
    job: Job[TDeps, TJobResult],
    *,
    store_history: bool = True,
    include_agent_tools: bool = True,
) -> list[AgentResponse[TJobResult]]

Execute a job across all team members in parallel.

Parameters:

Name	Type	Description	Default
job	Job[TDeps, TJobResult]	Job configuration to execute	required
store_history	bool	Whether to add job execution to conversation history	True
include_agent_tools	bool	Whether to include agent's tools alongside job tools	True

Returns:

Type	Description
list[AgentResponse[TJobResult]]	List of responses from all agents

Raises:

Type	Description
JobError	If job execution fails for any agent
ValueError	If job configuration is invalid

Source code in src/llmling_agent/delegation/team.py

async def run_job[TJobResult](
    self,
    job: Job[TDeps, TJobResult],
    *,
    store_history: bool = True,
    include_agent_tools: bool = True,
) -> list[AgentResponse[TJobResult]]:
    """Execute a job across all team members in parallel.

    Args:
        job: Job configuration to execute
        store_history: Whether to add job execution to conversation history
        include_agent_tools: Whether to include agent's tools alongside job tools

    Returns:
        List of responses from all agents

    Raises:
        JobError: If job execution fails for any agent
        ValueError: If job configuration is invalid
    """
    from llmling_agent.agent import Agent, StructuredAgent
    from llmling_agent.tasks import JobError

    responses: list[AgentResponse[TJobResult]] = []
    errors: dict[str, Exception] = {}
    start_time = get_now()

    # Validate dependencies for all agents
    if job.required_dependency is not None:
        invalid_agents = [
            agent.name
            for agent in self.iter_agents()
            if not isinstance(agent.context.data, job.required_dependency)
        ]
        if invalid_agents:
            msg = (
                f"Agents {', '.join(invalid_agents)} don't have required "
                f"dependency type: {job.required_dependency}"
            )
            raise JobError(msg)

    try:
        # Load knowledge for all agents if provided
        if job.knowledge:
            # TODO: resources
            tools = [t.name for t in job.get_tools()]
            await self.distribute(content="", tools=tools)

        prompt = await job.get_prompt()

        async def _run(agent: MessageNode[TDeps, TJobResult]):
            assert isinstance(agent, Agent | StructuredAgent)
            try:
                with agent.tools.temporary_tools(
                    job.get_tools(), exclusive=not include_agent_tools
                ):
                    start = perf_counter()
                    resp = AgentResponse(
                        agent_name=agent.name,
                        message=await agent.run(prompt, store_history=store_history),  # pyright: ignore
                        timing=perf_counter() - start,
                    )
                    responses.append(resp)
            except Exception as e:  # noqa: BLE001
                errors[agent.name] = e

        # Run job in parallel on all agents
        await asyncio.gather(*[_run(node) for node in self.agents])

        return TeamResponse(responses=responses, start_time=start_time, errors=errors)

    except Exception as e:
        msg = "Job execution failed"
        logger.exception(msg)
        raise JobError(msg) from e

TeamRun

Bases: BaseTeam[TDeps, TResult]

Handles team operations with monitoring.

Source code in src/llmling_agent/delegation/teamrun.py

class TeamRun[TDeps, TResult](BaseTeam[TDeps, TResult]):
    """Handles team operations with monitoring."""

    def __init__(
        self,
        agents: Sequence[MessageNode[TDeps, Any]],
        *,
        name: str | None = None,
        description: str | None = None,
        shared_prompt: str | None = None,
        validator: MessageNode[Any, TResult] | None = None,
        picker: AnyAgent[Any, Any] | None = None,
        num_picks: int | None = None,
        pick_prompt: str | None = None,
        # result_mode: ResultMode = "last",
    ):
        super().__init__(
            agents,
            name=name,
            description=description,
            shared_prompt=shared_prompt,
            picker=picker,
            num_picks=num_picks,
            pick_prompt=pick_prompt,
        )
        self.validator = validator
        self.result_mode = "last"

    def __prompt__(self) -> str:
        """Format team info for prompts."""
        members = " -> ".join(a.name for a in self.agents)
        desc = f" - {self.description}" if self.description else ""
        return f"Sequential Team '{self.name}'{desc}\nPipeline: {members}"

    async def _run(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        wait_for_connections: bool | None = None,
        message_id: str | None = None,
        conversation_id: str | None = None,
        **kwargs: Any,
    ) -> ChatMessage[TResult]:
        """Run agents sequentially and return combined message.

        This message wraps execute and extracts the ChatMessage in order to fulfill
        the "message protocol".
        """
        message_id = message_id or str(uuid4())

        result = await self.execute(*prompts, **kwargs)
        all_messages = [r.message for r in result if r.message]
        assert all_messages, "Error during execution, returned None for TeamRun"
        # Determine content based on mode
        match self.result_mode:
            case "last":
                content = all_messages[-1].content
            # case "concat":
            #     content = "\n".join(msg.format() for msg in all_messages)
            case _:
                msg = f"Invalid result mode: {self.result_mode}"
                raise ValueError(msg)

        return ChatMessage(
            content=content,
            role="assistant",
            name=self.name,
            associated_messages=all_messages,
            message_id=message_id,
            conversation_id=conversation_id,
            metadata={
                "execution_order": [r.agent_name for r in result],
                "start_time": result.start_time.isoformat(),
                "errors": {name: str(error) for name, error in result.errors.items()},
            },
        )

    async def execute(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        **kwargs: Any,
    ) -> TeamResponse[TResult]:
        """Start execution with optional monitoring."""
        self._team_talk.clear()
        start_time = get_now()
        final_prompt = list(prompts)
        if self.shared_prompt:
            final_prompt.insert(0, self.shared_prompt)

        responses = [
            i
            async for i in self.execute_iter(*final_prompt)
            if isinstance(i, AgentResponse)
        ]
        return TeamResponse(responses, start_time)

    async def run_iter(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        **kwargs: Any,
    ) -> AsyncIterator[ChatMessage[Any]]:
        """Yield messages from the execution chain."""
        async for item in self.execute_iter(*prompts, **kwargs):
            match item:
                case AgentResponse():
                    if item.message:
                        yield item.message
                case Talk():
                    pass

    async def execute_iter(
        self,
        *prompt: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        **kwargs: Any,
    ) -> AsyncIterator[Talk[Any] | AgentResponse[Any]]:
        from toprompt import to_prompt

        connections: list[Talk[Any]] = []
        try:
            combined_prompt = "\n".join([await to_prompt(p) for p in prompt])
            all_nodes = list(await self.pick_agents(combined_prompt))
            if self.validator:
                all_nodes.append(self.validator)
            first = all_nodes[0]
            connections = [
                source.connect_to(target, queued=True)
                for source, target in pairwise(all_nodes)
            ]
            for conn in connections:
                self._team_talk.append(conn)

            # First agent
            start = perf_counter()
            message = await first.run(*prompt, **kwargs)
            timing = perf_counter() - start
            response = AgentResponse[Any](first.name, message=message, timing=timing)
            yield response

            # Process through chain
            for connection in connections:
                target = connection.targets[0]
                target_name = target.name
                yield connection

                # Let errors propagate - they break the chain
                start = perf_counter()
                messages = await connection.trigger()

                # If this is the last node
                if target == all_nodes[-1]:
                    last_talk = Talk[Any](target, [], connection_type="run")
                    if response.message:
                        last_talk.stats.messages.append(response.message)
                    self._team_talk.append(last_talk)

                timing = perf_counter() - start
                msg = messages[0]
                response = AgentResponse[Any](target_name, message=msg, timing=timing)
                yield response

        finally:
            # Always clean up connections
            for connection in connections:
                connection.disconnect()

    @asynccontextmanager
    async def chain_stream(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
        require_all: bool = True,
        **kwargs: Any,
    ) -> AsyncIterator[StreamingResponseProtocol]:
        """Stream results through chain of team members."""
        from llmling_agent.agent import Agent, StructuredAgent
        from llmling_agent.delegation import TeamRun
        from llmling_agent_providers.base import StreamingResponseProtocol

        async with AsyncExitStack() as stack:
            streams: list[StreamingResponseProtocol[str]] = []
            current_message = prompts

            # Set up all streams
            for agent in self.agents:
                try:
                    assert isinstance(agent, TeamRun | Agent | StructuredAgent), (
                        "Cannot stream teams!"
                    )
                    stream = await stack.enter_async_context(
                        agent.run_stream(*current_message, **kwargs)
                    )
                    streams.append(stream)  # type: ignore
                    # Wait for complete response for next agent
                    async for chunk in stream.stream():
                        current_message = chunk
                        if stream.is_complete:
                            current_message = (stream.formatted_content,)  # type: ignore
                            break
                except Exception as e:
                    if require_all:
                        msg = f"Chain broken at {agent.name}: {e}"
                        raise ValueError(msg) from e
                    logger.warning("Chain handler %s failed: %s", agent.name, e)

            # Create a stream-like interface for the chain
            class ChainStream(StreamingResponseProtocol[str]):
                def __init__(self):
                    self.streams = streams
                    self.current_stream_idx = 0
                    self.is_complete = False
                    self.model_name = None

                def usage(self) -> Usage:
                    @dataclass
                    class Usage:
                        total_tokens: int | None
                        request_tokens: int | None
                        response_tokens: int | None

                    return Usage(0, 0, 0)

                async def stream(self) -> AsyncGenerator[str, None]:  # type: ignore
                    for idx, stream in enumerate(self.streams):
                        self.current_stream_idx = idx
                        async for chunk in stream.stream():
                            yield chunk
                            if idx == len(self.streams) - 1 and stream.is_complete:
                                self.is_complete = True

                async def stream_text(
                    self,
                    delta: bool = False,
                ) -> AsyncGenerator[str, None]:
                    for idx, stream in enumerate(self.streams):
                        self.current_stream_idx = idx
                        async for chunk in stream.stream_text(delta=delta):
                            yield chunk
                            if idx == len(self.streams) - 1 and stream.is_complete:
                                self.is_complete = True

            yield ChainStream()

    @asynccontextmanager
    async def run_stream(
        self,
        *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str],
        **kwargs: Any,
    ) -> AsyncIterator[StreamingResponseProtocol[TResult]]:
        """Stream responses through the chain.

        Provides same interface as Agent.run_stream.
        """
        async with self.chain_stream(*prompts, **kwargs) as stream:
            yield stream

__prompt__

__prompt__() -> str

Format team info for prompts.

Source code in src/llmling_agent/delegation/teamrun.py

def __prompt__(self) -> str:
    """Format team info for prompts."""
    members = " -> ".join(a.name for a in self.agents)
    desc = f" - {self.description}" if self.description else ""
    return f"Sequential Team '{self.name}'{desc}\nPipeline: {members}"

chain_stream async

chain_stream(
    *prompts: AnyPromptType | Image | PathLike[str] | None,
    require_all: bool = True,
    **kwargs: Any,
) -> AsyncIterator[StreamingResponseProtocol]

Stream results through chain of team members.

Source code in src/llmling_agent/delegation/teamrun.py

@asynccontextmanager
async def chain_stream(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
    require_all: bool = True,
    **kwargs: Any,
) -> AsyncIterator[StreamingResponseProtocol]:
    """Stream results through chain of team members."""
    from llmling_agent.agent import Agent, StructuredAgent
    from llmling_agent.delegation import TeamRun
    from llmling_agent_providers.base import StreamingResponseProtocol

    async with AsyncExitStack() as stack:
        streams: list[StreamingResponseProtocol[str]] = []
        current_message = prompts

        # Set up all streams
        for agent in self.agents:
            try:
                assert isinstance(agent, TeamRun | Agent | StructuredAgent), (
                    "Cannot stream teams!"
                )
                stream = await stack.enter_async_context(
                    agent.run_stream(*current_message, **kwargs)
                )
                streams.append(stream)  # type: ignore
                # Wait for complete response for next agent
                async for chunk in stream.stream():
                    current_message = chunk
                    if stream.is_complete:
                        current_message = (stream.formatted_content,)  # type: ignore
                        break
            except Exception as e:
                if require_all:
                    msg = f"Chain broken at {agent.name}: {e}"
                    raise ValueError(msg) from e
                logger.warning("Chain handler %s failed: %s", agent.name, e)

        # Create a stream-like interface for the chain
        class ChainStream(StreamingResponseProtocol[str]):
            def __init__(self):
                self.streams = streams
                self.current_stream_idx = 0
                self.is_complete = False
                self.model_name = None

            def usage(self) -> Usage:
                @dataclass
                class Usage:
                    total_tokens: int | None
                    request_tokens: int | None
                    response_tokens: int | None

                return Usage(0, 0, 0)

            async def stream(self) -> AsyncGenerator[str, None]:  # type: ignore
                for idx, stream in enumerate(self.streams):
                    self.current_stream_idx = idx
                    async for chunk in stream.stream():
                        yield chunk
                        if idx == len(self.streams) - 1 and stream.is_complete:
                            self.is_complete = True

            async def stream_text(
                self,
                delta: bool = False,
            ) -> AsyncGenerator[str, None]:
                for idx, stream in enumerate(self.streams):
                    self.current_stream_idx = idx
                    async for chunk in stream.stream_text(delta=delta):
                        yield chunk
                        if idx == len(self.streams) - 1 and stream.is_complete:
                            self.is_complete = True

        yield ChainStream()

execute async

execute(
    *prompts: AnyPromptType | Image | PathLike[str] | None, **kwargs: Any
) -> TeamResponse[TResult]

Start execution with optional monitoring.

Source code in src/llmling_agent/delegation/teamrun.py

async def execute(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str] | None,
    **kwargs: Any,
) -> TeamResponse[TResult]:
    """Start execution with optional monitoring."""
    self._team_talk.clear()
    start_time = get_now()
    final_prompt = list(prompts)
    if self.shared_prompt:
        final_prompt.insert(0, self.shared_prompt)

    responses = [
        i
        async for i in self.execute_iter(*final_prompt)
        if isinstance(i, AgentResponse)
    ]
    return TeamResponse(responses, start_time)

run_iter async

run_iter(
    *prompts: AnyPromptType | Image | PathLike[str], **kwargs: Any
) -> AsyncIterator[ChatMessage[Any]]

Yield messages from the execution chain.

Source code in src/llmling_agent/delegation/teamrun.py

async def run_iter(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    **kwargs: Any,
) -> AsyncIterator[ChatMessage[Any]]:
    """Yield messages from the execution chain."""
    async for item in self.execute_iter(*prompts, **kwargs):
        match item:
            case AgentResponse():
                if item.message:
                    yield item.message
            case Talk():
                pass

run_stream async

run_stream(
    *prompts: AnyPromptType | Image | PathLike[str], **kwargs: Any
) -> AsyncIterator[StreamingResponseProtocol[TResult]]

Stream responses through the chain.

Provides same interface as Agent.run_stream.

Source code in src/llmling_agent/delegation/teamrun.py

@asynccontextmanager
async def run_stream(
    self,
    *prompts: AnyPromptType | PIL.Image.Image | os.PathLike[str],
    **kwargs: Any,
) -> AsyncIterator[StreamingResponseProtocol[TResult]]:
    """Stream responses through the chain.

    Provides same interface as Agent.run_stream.
    """
    async with self.chain_stream(*prompts, **kwargs) as stream:
        yield stream

Tool dataclass

Information about a registered tool.

Source code in src/llmling_agent/tools/base.py

@dataclass
class Tool:
    """Information about a registered tool."""

    callable: LLMCallableTool
    """The actual tool implementation"""

    enabled: bool = True
    """Whether the tool is currently enabled"""

    source: ToolSource = "runtime"
    """Where the tool came from."""

    priority: int = 100
    """Priority for tool execution (lower = higher priority)"""

    requires_confirmation: bool = False
    """Whether tool execution needs explicit confirmation"""

    requires_capability: str | None = None
    """Optional capability required to use this tool"""

    agent_name: str | None = None
    """The agent name as an identifier for agent-as-a-tool."""

    metadata: dict[str, str] = field(default_factory=dict)
    """Additional tool metadata"""

    cache_enabled: bool = False
    """Whether to enable caching for this tool."""

    @property
    def schema(self) -> py2openai.OpenAIFunctionTool:
        """Get the OpenAI function schema for the tool."""
        return self.callable.get_schema()

    @property
    def name(self) -> str:
        """Get tool name."""
        return self.callable.name

    @property
    def description(self) -> str | None:
        """Get tool description."""
        return self.callable.description

    def matches_filter(self, state: Literal["all", "enabled", "disabled"]) -> bool:
        """Check if tool matches state filter."""
        match state:
            case "all":
                return True
            case "enabled":
                return self.enabled
            case "disabled":
                return not self.enabled

    @property
    def parameters(self) -> list[ToolParameter]:
        """Get information about tool parameters."""
        schema = self.schema["function"]
        properties: dict[str, Property] = schema.get("properties", {})  # type: ignore
        required: list[str] = schema.get("required", [])  # type: ignore

        return [
            ToolParameter(
                name=name,
                required=name in required,
                type_info=details.get("type"),
                description=details.get("description"),
            )
            for name, details in properties.items()
        ]

    def format_info(self, indent: str = "  ") -> str:
        """Format complete tool information."""
        lines = [f"{indent}→ {self.name}"]
        if self.description:
            lines.append(f"{indent}  {self.description}")
        if self.parameters:
            lines.append(f"{indent}  Parameters:")
            lines.extend(f"{indent}    {param}" for param in self.parameters)
        if self.metadata:
            lines.append(f"{indent}  Metadata:")
            lines.extend(f"{indent}    {k}: {v}" for k, v in self.metadata.items())
        return "\n".join(lines)

    async def execute(self, *args: Any, **kwargs: Any) -> Any:
        """Execute tool, handling both sync and async cases."""
        fn = track_tool(self.name)(self.callable.callable)
        return await execute(fn, *args, **kwargs, use_thread=True)

    @classmethod
    def from_code(
        cls,
        code: str,
        name: str | None = None,
        description: str | None = None,
    ) -> Self:
        """Create a tool from a code string."""
        namespace: dict[str, Any] = {}
        exec(code, namespace)
        func = next((v for v in namespace.values() if callable(v)), None)
        if not func:
            msg = "No callable found in provided code"
            raise ValueError(msg)
        return cls.from_callable(
            func, name_override=name, description_override=description
        )

    @classmethod
    def from_callable(
        cls,
        fn: Callable[..., Any] | str,
        *,
        name_override: str | None = None,
        description_override: str | None = None,
        schema_override: py2openai.OpenAIFunctionDefinition | None = None,
        **kwargs: Any,
    ) -> Self:
        tool = LLMCallableTool.from_callable(
            fn,
            name_override=name_override,
            description_override=description_override,
            schema_override=schema_override,
        )
        return cls(tool, **kwargs)

    @classmethod
    def from_crewai_tool(
        cls,
        tool: Any,
        *,
        name_override: str | None = None,
        description_override: str | None = None,
        schema_override: py2openai.OpenAIFunctionDefinition | None = None,
        **kwargs: Any,
    ) -> Self:
        """Allows importing crewai tools."""
        # vaidate_import("crewai_tools", "crewai")
        try:
            from crewai.tools import BaseTool as CrewAiBaseTool
        except ImportError as e:
            msg = "crewai package not found. Please install it with 'pip install crewai'"
            raise ImportError(msg) from e

        if not isinstance(tool, CrewAiBaseTool):
            msg = f"Expected CrewAI BaseTool, got {type(tool)}"
            raise TypeError(msg)

        return cls.from_callable(
            tool._run,
            name_override=name_override or tool.__class__.__name__.removesuffix("Tool"),
            description_override=description_override or tool.description,
            schema_override=schema_override,
            **kwargs,
        )

    @classmethod
    def from_langchain_tool(
        cls,
        tool: Any,
        *,
        name_override: str | None = None,
        description_override: str | None = None,
        schema_override: py2openai.OpenAIFunctionDefinition | None = None,
        **kwargs: Any,
    ) -> Self:
        """Create a tool from a LangChain tool."""
        # vaidate_import("langchain_core", "langchain")
        try:
            from langchain_core.tools import BaseTool as LangChainBaseTool
        except ImportError as e:
            msg = "langchain-core package not found."
            raise ImportError(msg) from e

        if not isinstance(tool, LangChainBaseTool):
            msg = f"Expected LangChain BaseTool, got {type(tool)}"
            raise TypeError(msg)

        return cls.from_callable(
            tool.invoke,
            name_override=name_override or tool.name,
            description_override=description_override or tool.description,
            schema_override=schema_override,
            **kwargs,
        )

    @classmethod
    def from_autogen_tool(
        cls,
        tool: Any,
        *,
        name_override: str | None = None,
        description_override: str | None = None,
        schema_override: py2openai.OpenAIFunctionDefinition | None = None,
        **kwargs: Any,
    ) -> Self:
        """Create a tool from a AutoGen tool."""
        # vaidate_import("autogen_core", "autogen")
        try:
            from autogen_core import CancellationToken
            from autogen_core.tools import BaseTool
        except ImportError as e:
            msg = "autogent_core package not found."
            raise ImportError(msg) from e

        if not isinstance(tool, BaseTool):
            msg = f"Expected AutoGent BaseTool, got {type(tool)}"
            raise TypeError(msg)
        token = CancellationToken()

        input_model = tool.__class__.__orig_bases__[0].__args__[0]  # type: ignore

        name = name_override or tool.name or tool.__class__.__name__.removesuffix("Tool")
        description = (
            description_override
            or tool.description
            or inspect.getdoc(tool.__class__)
            or ""
        )

        async def wrapper(**kwargs: Any) -> Any:
            # Convert kwargs to the expected input model
            model = input_model(**kwargs)
            return await tool.run(model, cancellation_token=token)

        return cls.from_callable(
            wrapper,  # type: ignore
            name_override=name,
            description_override=description,
            schema_override=schema_override,
            **kwargs,
        )

agent_name class-attribute instance-attribute

agent_name: str | None = None

The agent name as an identifier for agent-as-a-tool.

cache_enabled class-attribute instance-attribute

cache_enabled: bool = False

Whether to enable caching for this tool.

callable instance-attribute

callable: LLMCallableTool

The actual tool implementation

description property

description: str | None

Get tool description.

enabled class-attribute instance-attribute

enabled: bool = True

Whether the tool is currently enabled

metadata class-attribute instance-attribute

metadata: dict[str, str] = field(default_factory=dict)

Additional tool metadata

name property

name: str

Get tool name.

parameters property

parameters: list[ToolParameter]

Get information about tool parameters.

priority class-attribute instance-attribute

priority: int = 100

Priority for tool execution (lower = higher priority)

requires_capability class-attribute instance-attribute

requires_capability: str | None = None

Optional capability required to use this tool

requires_confirmation class-attribute instance-attribute

requires_confirmation: bool = False

Whether tool execution needs explicit confirmation

schema property

schema: OpenAIFunctionTool

Get the OpenAI function schema for the tool.

source class-attribute instance-attribute

source: ToolSource = 'runtime'

Where the tool came from.

execute async

execute(*args: Any, **kwargs: Any) -> Any

Execute tool, handling both sync and async cases.

Source code in src/llmling_agent/tools/base.py

async def execute(self, *args: Any, **kwargs: Any) -> Any:
    """Execute tool, handling both sync and async cases."""
    fn = track_tool(self.name)(self.callable.callable)
    return await execute(fn, *args, **kwargs, use_thread=True)

format_info

format_info(indent: str = '  ') -> str

Format complete tool information.

Source code in src/llmling_agent/tools/base.py

def format_info(self, indent: str = "  ") -> str:
    """Format complete tool information."""
    lines = [f"{indent}→ {self.name}"]
    if self.description:
        lines.append(f"{indent}  {self.description}")
    if self.parameters:
        lines.append(f"{indent}  Parameters:")
        lines.extend(f"{indent}    {param}" for param in self.parameters)
    if self.metadata:
        lines.append(f"{indent}  Metadata:")
        lines.extend(f"{indent}    {k}: {v}" for k, v in self.metadata.items())
    return "\n".join(lines)

from_autogen_tool classmethod

from_autogen_tool(
    tool: Any,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    schema_override: OpenAIFunctionDefinition | None = None,
    **kwargs: Any,
) -> Self

Create a tool from a AutoGen tool.

Source code in src/llmling_agent/tools/base.py

@classmethod
def from_autogen_tool(
    cls,
    tool: Any,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    schema_override: py2openai.OpenAIFunctionDefinition | None = None,
    **kwargs: Any,
) -> Self:
    """Create a tool from a AutoGen tool."""
    # vaidate_import("autogen_core", "autogen")
    try:
        from autogen_core import CancellationToken
        from autogen_core.tools import BaseTool
    except ImportError as e:
        msg = "autogent_core package not found."
        raise ImportError(msg) from e

    if not isinstance(tool, BaseTool):
        msg = f"Expected AutoGent BaseTool, got {type(tool)}"
        raise TypeError(msg)
    token = CancellationToken()

    input_model = tool.__class__.__orig_bases__[0].__args__[0]  # type: ignore

    name = name_override or tool.name or tool.__class__.__name__.removesuffix("Tool")
    description = (
        description_override
        or tool.description
        or inspect.getdoc(tool.__class__)
        or ""
    )

    async def wrapper(**kwargs: Any) -> Any:
        # Convert kwargs to the expected input model
        model = input_model(**kwargs)
        return await tool.run(model, cancellation_token=token)

    return cls.from_callable(
        wrapper,  # type: ignore
        name_override=name,
        description_override=description,
        schema_override=schema_override,
        **kwargs,
    )

from_code classmethod

from_code(code: str, name: str | None = None, description: str | None = None) -> Self

Create a tool from a code string.

Source code in src/llmling_agent/tools/base.py

@classmethod
def from_code(
    cls,
    code: str,
    name: str | None = None,
    description: str | None = None,
) -> Self:
    """Create a tool from a code string."""
    namespace: dict[str, Any] = {}
    exec(code, namespace)
    func = next((v for v in namespace.values() if callable(v)), None)
    if not func:
        msg = "No callable found in provided code"
        raise ValueError(msg)
    return cls.from_callable(
        func, name_override=name, description_override=description
    )

from_crewai_tool classmethod

from_crewai_tool(
    tool: Any,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    schema_override: OpenAIFunctionDefinition | None = None,
    **kwargs: Any,
) -> Self

Allows importing crewai tools.

Source code in src/llmling_agent/tools/base.py

@classmethod
def from_crewai_tool(
    cls,
    tool: Any,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    schema_override: py2openai.OpenAIFunctionDefinition | None = None,
    **kwargs: Any,
) -> Self:
    """Allows importing crewai tools."""
    # vaidate_import("crewai_tools", "crewai")
    try:
        from crewai.tools import BaseTool as CrewAiBaseTool
    except ImportError as e:
        msg = "crewai package not found. Please install it with 'pip install crewai'"
        raise ImportError(msg) from e

    if not isinstance(tool, CrewAiBaseTool):
        msg = f"Expected CrewAI BaseTool, got {type(tool)}"
        raise TypeError(msg)

    return cls.from_callable(
        tool._run,
        name_override=name_override or tool.__class__.__name__.removesuffix("Tool"),
        description_override=description_override or tool.description,
        schema_override=schema_override,
        **kwargs,
    )

from_langchain_tool classmethod

from_langchain_tool(
    tool: Any,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    schema_override: OpenAIFunctionDefinition | None = None,
    **kwargs: Any,
) -> Self

Create a tool from a LangChain tool.

Source code in src/llmling_agent/tools/base.py

@classmethod
def from_langchain_tool(
    cls,
    tool: Any,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    schema_override: py2openai.OpenAIFunctionDefinition | None = None,
    **kwargs: Any,
) -> Self:
    """Create a tool from a LangChain tool."""
    # vaidate_import("langchain_core", "langchain")
    try:
        from langchain_core.tools import BaseTool as LangChainBaseTool
    except ImportError as e:
        msg = "langchain-core package not found."
        raise ImportError(msg) from e

    if not isinstance(tool, LangChainBaseTool):
        msg = f"Expected LangChain BaseTool, got {type(tool)}"
        raise TypeError(msg)

    return cls.from_callable(
        tool.invoke,
        name_override=name_override or tool.name,
        description_override=description_override or tool.description,
        schema_override=schema_override,
        **kwargs,
    )

matches_filter

matches_filter(state: Literal['all', 'enabled', 'disabled']) -> bool

Check if tool matches state filter.

Source code in src/llmling_agent/tools/base.py

def matches_filter(self, state: Literal["all", "enabled", "disabled"]) -> bool:
    """Check if tool matches state filter."""
    match state:
        case "all":
            return True
        case "enabled":
            return self.enabled
        case "disabled":
            return not self.enabled

ToolCallInfo

Bases: Schema

Information about an executed tool call.

Source code in src/llmling_agent/tools/tool_call_info.py

class ToolCallInfo(Schema):
    """Information about an executed tool call."""

    tool_name: str
    """Name of the tool that was called."""

    args: dict[str, Any]
    """Arguments passed to the tool."""

    result: Any
    """Result returned by the tool."""

    agent_name: str
    """Name of the calling agent."""

    tool_call_id: str = Field(default_factory=lambda: str(uuid4()))
    """ID provided by the model (e.g. OpenAI function call ID)."""

    timestamp: datetime = Field(default_factory=get_now)
    """When the tool was called."""

    message_id: str | None = None
    """ID of the message that triggered this tool call."""

    context_data: Any | None = None
    """Optional context data that was passed to the agent's run() method."""

    error: str | None = None
    """Error message if the tool call failed."""

    timing: float | None = None
    """Time taken for this specific tool call in seconds."""

    agent_tool_name: str | None = None
    """If this tool is agent-based, the name of that agent."""

    def format(
        self,
        style: FormatStyle = "simple",
        *,
        template: str | None = None,
        variables: dict[str, Any] | None = None,
        show_timing: bool = True,
        show_ids: bool = False,
    ) -> str:
        """Format tool call information with configurable style.

        Args:
            style: Predefined style to use:
                - simple: Compact single-line format
                - detailed: Multi-line with all details
                - markdown: Formatted markdown with syntax highlighting
            template: Optional custom template (required if style="custom")
            variables: Additional variables for template rendering
            show_timing: Whether to include execution timing
            show_ids: Whether to include tool_call_id and message_id

        Returns:
            Formatted tool call information

        Raises:
            ValueError: If style is invalid or custom template is missing
        """
        from jinjarope import Environment

        # Select template
        if template:
            template_str = template
        elif style in TEMPLATES:
            template_str = TEMPLATES[style]
        else:
            msg = f"Invalid style: {style}"
            raise ValueError(msg)

        # Prepare template variables
        vars_ = {
            "tool_name": self.tool_name,
            "args": self.args,  # No pre-formatting needed
            "result": self.result,
            "error": self.error,
            "agent_name": self.agent_name,
            "timestamp": self.timestamp,
            "timing": self.timing if show_timing else None,
            "agent_tool_name": self.agent_tool_name,
        }

        if show_ids:
            vars_.update({
                "tool_call_id": self.tool_call_id,
                "message_id": self.message_id,
            })

        if variables:
            vars_.update(variables)

        # Render template
        env = Environment(trim_blocks=True, lstrip_blocks=True)
        env.filters["repr"] = repr  # Add repr filter
        template_obj = env.from_string(template_str)
        return template_obj.render(**vars_)

agent_name instance-attribute

agent_name: str

Name of the calling agent.

agent_tool_name class-attribute instance-attribute

agent_tool_name: str | None = None

If this tool is agent-based, the name of that agent.

args instance-attribute

args: dict[str, Any]

Arguments passed to the tool.

context_data class-attribute instance-attribute

context_data: Any | None = None

Optional context data that was passed to the agent's run() method.

error class-attribute instance-attribute

error: str | None = None

Error message if the tool call failed.

message_id class-attribute instance-attribute

message_id: str | None = None

ID of the message that triggered this tool call.

result instance-attribute

result: Any

Result returned by the tool.

timestamp class-attribute instance-attribute

timestamp: datetime = Field(default_factory=get_now)

When the tool was called.

timing class-attribute instance-attribute

timing: float | None = None

Time taken for this specific tool call in seconds.

tool_call_id class-attribute instance-attribute

tool_call_id: str = Field(default_factory=lambda: str(uuid4()))

ID provided by the model (e.g. OpenAI function call ID).

tool_name instance-attribute

tool_name: str

Name of the tool that was called.

format

format(
    style: FormatStyle = "simple",
    *,
    template: str | None = None,
    variables: dict[str, Any] | None = None,
    show_timing: bool = True,
    show_ids: bool = False,
) -> str

Format tool call information with configurable style.

Parameters:

Name	Type	Description	Default
style	FormatStyle	Predefined style to use: - simple: Compact single-line format - detailed: Multi-line with all details - markdown: Formatted markdown with syntax highlighting	'simple'
template	str | None	Optional custom template (required if style="custom")	None
variables	dict[str, Any] | None	Additional variables for template rendering	None
show_timing	bool	Whether to include execution timing	True
show_ids	bool	Whether to include tool_call_id and message_id	False

Returns:

Type	Description
str	Formatted tool call information

Raises:

Type	Description
ValueError	If style is invalid or custom template is missing

Source code in src/llmling_agent/tools/tool_call_info.py

def format(
    self,
    style: FormatStyle = "simple",
    *,
    template: str | None = None,
    variables: dict[str, Any] | None = None,
    show_timing: bool = True,
    show_ids: bool = False,
) -> str:
    """Format tool call information with configurable style.

    Args:
        style: Predefined style to use:
            - simple: Compact single-line format
            - detailed: Multi-line with all details
            - markdown: Formatted markdown with syntax highlighting
        template: Optional custom template (required if style="custom")
        variables: Additional variables for template rendering
        show_timing: Whether to include execution timing
        show_ids: Whether to include tool_call_id and message_id

    Returns:
        Formatted tool call information

    Raises:
        ValueError: If style is invalid or custom template is missing
    """
    from jinjarope import Environment

    # Select template
    if template:
        template_str = template
    elif style in TEMPLATES:
        template_str = TEMPLATES[style]
    else:
        msg = f"Invalid style: {style}"
        raise ValueError(msg)

    # Prepare template variables
    vars_ = {
        "tool_name": self.tool_name,
        "args": self.args,  # No pre-formatting needed
        "result": self.result,
        "error": self.error,
        "agent_name": self.agent_name,
        "timestamp": self.timestamp,
        "timing": self.timing if show_timing else None,
        "agent_tool_name": self.agent_tool_name,
    }

    if show_ids:
        vars_.update({
            "tool_call_id": self.tool_call_id,
            "message_id": self.message_id,
        })

    if variables:
        vars_.update(variables)

    # Render template
    env = Environment(trim_blocks=True, lstrip_blocks=True)
    env.filters["repr"] = repr  # Add repr filter
    template_obj = env.from_string(template_str)
    return template_obj.render(**vars_)

VideoURLContent

Bases: VideoContent

Video from URL.

Source code in src/llmling_agent/models/content.py

class VideoURLContent(VideoContent):
    """Video from URL."""

    type: Literal["video_url"] = Field("video_url", init=False)
    """URL-based video."""

    url: str
    """URL to the video."""

    def to_openai_format(self) -> dict[str, Any]:
        """Convert to OpenAI API format for video models."""
        content = {"url": self.url, "format": self.format or "auto"}
        return {"type": "video", "video": content}

type class-attribute instance-attribute

type: Literal['video_url'] = Field('video_url', init=False)

URL-based video.

url instance-attribute

url: str

URL to the video.

to_openai_format

to_openai_format() -> dict[str, Any]

Convert to OpenAI API format for video models.

Source code in src/llmling_agent/models/content.py

def to_openai_format(self) -> dict[str, Any]:
    """Convert to OpenAI API format for video models."""
    content = {"url": self.url, "format": self.format or "auto"}
    return {"type": "video", "video": content}