Welcome to LLMling-Agent

Easy Agent Configuration

LLMling-agent excels at static YAML-based agent configuration:

  • Define agents in not-seen-yet detail in pure YAML (pydantic-backed)
  • Expansive YAML schema for linters for providing maxium help setting up YAML files
  • Agent "connection" setup via YAML allowing simple workflows without having to use a step-based appraoch
  • Result type definitions via YAML (StructuredAgents)
  • Configuration inheritance and reuse
  • Environment/tools/resource configuration
  • Extensive schema validation and linting
  • Type-safe structured responses

True Async Framework

An async-first Agent framwork. Unlike most other frameworks, where async is only done at a workflow-step level, this framework is async-first in its design.

Built for modern async Python from the ground up.

  • Proper async context management
  • Non-blocking operations
  • Streaming responses
  • Resource cleanup
  • Background task handling
  • Parallel initialization
  • Automatic resource management

Type-Safety on Pydantic-Level

  • Excellently typed user APIs
  • A lot of love in detail to provide maximum type safety
  • Type-safe message passing
  • Type-safe Agent-Team forming
  • Type-safe task execution
  • Structured response handling
  • Flexible routing options
  • Cost and token tracking

Pool-Based Architecture

Central coordination point for multi-agent systems:

  • Type-safe dependency injection
  • Shared resource management
  • Session and history management
  • Dynamic agent / team creation/cloning
  • Agent discovery and access control
  • Central monitoring and statistics
  • Common storage configuration

Provider Architecture

Flexible provider system decoupling agent logic from AI implementation. Humans, LLM libraries and Callables can all be used as the "Agent" brain and be switched flawlessly, allowing unique AI-Human interactions.

  • Modular provider architecture
  • First-class Pydantic-AI support (recommended)
  • LiteLLM integration for model variety
  • Human-in-the-loop provider
  • Custom provider support
  • Callable provider
  • UI/Observer integration
  • Consistent interface across providers
  • Provider-specific optimizations
  • Stream support
  • Multi-modal support for LitemLLM (experimental)

Teams and Execution

Flexible team operations and execution patterns:

  • Multiple execution modes (parallel/sequential)
  • Rich monitoring capabilities
  • Connection mechanisms (>>, &, |)
  • Type-safe team creation
  • Execution statistics and cost tracking
  • Background execution with monitoring

MCP server integration

MCP server support for Agents.

  • Agents with MCP server configurations will "connect" on async initialization
  • MCP severs can provide agent tools
  • More extensive coverage of the MCP servers soon.

Event System

React to changes and automate workflows:

  • File system monitoring with patterns
  • Webhook endpoints (coming soon)
  • Configurable event debouncing
  • Event filtering and routing
  • Custom trigger types
  • Automatic context loading
  • Event-driven agent execution

Command System

Rich command system across all interfaces:

  • Unified command system
  • Tool management commands
  • Resource inspection
  • Model configuration
  • History access and search
  • Runtime configuration
  • Session management
  • Cross-interface consistency

Database & Logging

Comprehensive interaction tracking:

  • Multiple storage providers
  • Configurable logging levels
  • Message history tracking
  • Maximum flexibility in formatting conversations
  • Tool usage monitoring
  • Cost and token tracking
  • Query capabilities
  • Session recovery
  • Flexible storage backends
  • Sophisticated conversation filtering for agents

Interaction Patterns

Rich patterns for agent interaction and collaboration:

  • Automatic and human-in-the-loop decisions
  • Concurrent agent execution
  • Non-blocking forwarding mechanisms
  • Blocking when needed (await_response)
  • Automatic result routing (>>)
  • Team formation helper (pick)
  • Message broadcasting
  • Priority-based routing
  • Delayed forwarding
  • Connection monitoring

Multi-modal support

Both LiteLLM and Pydantic-AI (using a workaround) support Image input. Additional content types will be integrated soon!

Multiple Interfaces

Multiple ways to interact:

  • Rich CLI interface using prompt-toolkit with expansive command system
  • Web UI with Gradio
  • Terminal UI with Textual (WIP)
  • Python API
  • Consistent experience
  • Cross-interface features
  • Unified command system

Task System

Independent work definitions with type safety:

  • Tasks define requirements, not implementations
  • Type-safe dependency requirements
  • Automatic tool provisioning
  • Knowledge source integration
  • Runtime validation
  • Reusable task definitions
  • YAML-based task configuration

Capability System

Fine-grained permission control for agents:

  • Automatic tool availability based on capabilities
  • Granular access level management
  • Resource access control
  • Tool usage permissions
  • History and stats access levels
  • Agent interaction capabilites

Knowledge Management

Comprehensive knowledge integration:

  • Multiple knowledge sources
  • Rich resource handling
  • Dynamic prompt integration
  • Automatic context loading
  • Markdown conversion
  • Token-aware context management
  • Parallel resource initialization