Multi-Agent Orchestration Patterns¶

LLMling-agent offers several patterns for coordinating multiple agents. Each pattern has its strengths and is suited for different use cases.

Quick Reference¶

Pattern	Use Case	Complexity	Type Safety	Visualization	Control Flow
Direct Connections	Simple A → B flows	Low	✅	❌	Automatic
Controlled Interactions	Interactive/supervised flows	Low	✅	❌	Manual/Interactive
Decision Making (pick)	Agent selection/routing	Low	✅	❌	Agent-driven
Agent as Tool	Hierarchical/expert	Low	✅	❌	Parent-driven
Teams	Parallel/group ops	Medium	✅	❌	Coordinated
Decorator Pattern	Testing/scripted	Medium	✅	❌	Programmatic
Pydantic Graph	Complex workflows	High	✅	✅	Graph-based

1. Direct Connections (Simple Forwarding)¶

Best for: Simple linear flows between agents

analyzer = Agent("analyzer")
planner = Agent("planner")
executor = Agent("executor")

# Chain connections
analyzer >> planner >> executor

# Or with configuration
analyzer.connect_to(planner, connection_type="run")

2. Agent as Tool (Hierarchical)¶

Best for: Expert consultation patterns where one agent calls others as needed

main_agent = Agent("coordinator")
expert = Agent("expert")

# Register expert as tool
main_agent.register_worker(
    expert,
    name="consult_expert",
    reset_history_on_run=True
)

3. Teams (Group Operations)¶

Best for: Parallel execution or group coordination

team = Team([agent1, agent2, agent3])
result = await team.run_parallel("Analyze this data")

# Or chain through team
team.chain("Process this sequentially")

4. Decorator Pattern (Testing/Scripting)¶

Best for: Scripted interactions and testing flows

@with_nodes("analyzer", "planner")
async def analysis_flow(analyzer: Agent, planner: Agent):
    result = await analyzer.run("Analyze this")
    return await planner.run(result.content)