Supervisor Graph with LangGraph

graph.py

python

1from typing import Literal, TypedDict
2
3from langgraph.graph import END, StateGraph
4
5
6class State(TypedDict):
7    question: str
8    context: list[str]
9    answer: str
10    next: Literal["retrieve", "answer", "done"]
11
12
13def supervisor(state: State) -> State:
14    nxt = "retrieve" if not state["context"] else "answer"
15    return {**state, "next": nxt}
16
17
18def retrieve(state: State) -> State:
19    # Replace with hybrid retrieval over your corpus.
20    return {**state, "context": ["Chunk the corpus by structure, not size."]}
21
22
23def answer(state: State) -> State:
24    return {**state, "answer": " ".join(state["context"]), "next": "done"}
25
26
27graph = StateGraph(State)
28graph.add_node("supervisor", supervisor)
29graph.add_node("retrieve", retrieve)
30graph.add_node("answer", answer)
31graph.set_entry_point("supervisor")
32graph.add_conditional_edges(
33    "supervisor", lambda s: s["next"], {"retrieve": "retrieve", "answer": "answer"}
34)
35graph.add_edge("retrieve", "supervisor")
36graph.add_edge("answer", END)
37
38app = graph.compile()
39print(app.invoke({"question": "How should I chunk documents?", "context": [], "answer": "", "next": "retrieve"}))

Go further

Full example on GitHubgithub.com/stackunseen/examples/tree/main/langgraph-supervisor Framework docsOfficial documentation On the mapAI Application → Agent

Run it

From a clone of stackunseen/examples:

bash

git clone https://github.com/stackunseen/examples
cd examples
cd langgraph-supervisor
pip install langgraph
python graph.py

Why a graph

When the routing logic matters more than the prompts, an explicit graph makes the control flow reviewable. Every edge is visible in code and in the trace.

What to notice

The supervisor node is deterministic here. In practice it is often a small model call, but keeping it deterministic for routing decisions you can express in code saves cost and makes tests stable.