Skip to content

multi-agent.md

Latest commit

 

History

History
41 lines (29 loc) · 1.82 KB

multi-agent.md

File metadata and controls

41 lines (29 loc) · 1.82 KB
Raw

Multi-agent systems

Audience: authors composing specialists into a durable Conductor-agent graph.

Prerequisites

Each participating agent needs a unique stable name and bounded execution policy. Start with a single agent and a tested tool before introducing delegation.

Agent(agents=[...], strategy=...) supports SEQUENTIAL, PARALLEL, HANDOFF, ROUTER, ROUND_ROBIN, RANDOM, SWARM, MANUAL, and PLAN_EXECUTE.

  • Sequential and parallel strategies compose deterministic work.
  • Handoff, router, and swarm transfer control between agents; restrict transitions and define an empty-safe fallback.
  • PLAN_EXECUTE compiles a typed Plan into a durable sub-workflow and can replan.
  • agent_tool(child) calls a child agent as a tool instead of transferring control.

Set a termination condition and a turn/token limit for every open-ended design. See termination, plans, and writing examples.

Handoffs, routers, and plans

Use HANDOFF when the model should select a specialist, ROUTER when an explicit router chooses one, and agent_tool(child) when a parent needs a child's result without transferring control. PLAN_EXECUTE is for typed, inspectable plans that must survive restarts; provide a planner and an optional fallback/replan policy. Restrict allowed transitions so an unexpected model output cannot reach an unsafe specialist.

Expected result and failures

The compiled parent contains durable child/sub-workflow work and each child is visible in execution history. If a graph loops, add max_turns, termination, and a bounded planner/replan policy before retrying.

Next steps

Read termination, stateful agents, and the agent definition reference.