Creativity · Agent Protocol
LangGraph Supervisor Pattern
LangGraph, LangChain's graph-based agent framework, has an idiomatic pattern called 'supervisor'. A supervisor node routes the current message to one of several named worker agents, each a subgraph with its own tools and prompts. The supervisor reads the accumulated state, decides whom to run next, and loops until a termination condition. It's orchestrator-worker encoded as a state graph rather than an ad-hoc orchestrator prompt.
Protocol facts
- Sponsor
- LangChain
- Status
- stable
- Spec
- https://langchain-ai.github.io/langgraph/tutorials/multi_agent/agent_supervisor/
- Interop with
- LangChain, A2A, OpenAI, Anthropic, Tavily
Frequently asked questions
Why a state graph instead of a free-form orchestrator prompt?
The graph makes control flow explicit — you can see the transitions, checkpoint state between nodes, replay runs, and inspect exactly which worker ran when. LangGraph trades flexibility for observability.
How does the supervisor pick a worker?
Usually a structured-output call: the supervisor LLM sees the current state and returns the name of the next worker — or 'FINISH' to stop. It's a controlled router on top of a bigger graph.
How does this relate to A2A?
LangGraph supervisors keep workers in the same process. A2A is how you swap an in-process worker for a remote worker owned by a different team, without rewriting the supervisor.
Sources
- LangGraph — Agent supervisor tutorial — accessed 2026-04-20