LangGraph Supervisor Research Desk

What This Builds

This recipe implements the supervisor multi-agent architecture: a central supervisor agent coordinates several specialized worker agents. The supervisor controls communication flow and delegation, deciding which agent to call next based on the conversation and routing control back to itself after each worker finishes.

The concrete use case is a small research desk. A supervisor routes a user question to:

• a research agent that searches the web and records notes, and
• an analysis agent that does calculations or structured reasoning over the gathered facts.

The supervisor keeps delegating until it has enough to answer, then returns the final response. Optionally, a vector store gives the research agent retrieval over a private corpus.

The Stack

• LangGraph + langgraph-supervisor — a prebuilt library for hierarchical multi-agent systems. You create worker agents with create_react_agent, then wire them under a create_supervisor(...) graph that handles routing and handoffs.
• LangChain model bindings — Claude (via the Anthropic Startup Program) or any chat model through OpenRouter for the supervisor and workers; smaller/cheaper models for workers, a stronger model for the supervisor.
• Qdrant (Cloud free tier) — optional vector store backing the research agent’s retrieval tool so it can ground answers in your own documents alongside web search.
• LangSmith — optional tracing to watch routing decisions and worker handoffs step by step.

Step-by-Step Outline

1. Build worker agents. Define research_agent (with a web-search tool and, optionally, a Qdrant-backed retriever) and analysis_agent (with a calculator/structured-reasoning tool), each via create_react_agent with a clear name and instructions.
2. Create the supervisor. Pass the workers into create_supervisor([...], model=..., prompt=...) describing when to use each agent and the stopping condition.
3. Compile and run. Compile the supervisor graph and invoke it with the user’s question; the supervisor decides the order of worker calls.
4. Inspect routing. Use LangSmith (or stream graph events) to confirm the supervisor isn’t over-delegating or looping.
5. Extend. Add more specialists (a writer, a critic) or nest supervisors so a top-level planner delegates to mid-level supervisors that each control a group of domain agents.

Why This Shape Works

Routing through a single supervisor gives you separation of concerns — each worker has a focused prompt and toolset — while keeping one place responsible for “what next?” and “are we done?”. This is the same hierarchical structure that scales to networks of agents, and langgraph-supervisor removes the boilerplate of wiring handoffs and shared state by hand.

Source

LangChain, LangGraph Multi-Agent Supervisor reference and tutorial (langgraph-supervisor): https://reference.langchain.com/python/langgraph-supervisor

LangChain docs, Build a personal assistant with subagents (supervisor pattern overview): https://docs.langchain.com/oss/python/langchain/multi-agent/subagents-personal-assistant