How do you model roles (planner, researcher, writer, critic) as nodes/agents in LangGraph?

What you need to know

Supervisor as code

1from typing import Literal2from langgraph.types import Command34def supervisor(state) -> Command[Literal["researcher", "writer", "critic", "__end__"]]:5    if not state.get("findings"):6        return Command(goto="researcher")7    if not state.get("draft"):8        return Command(goto="writer")9    if state.get("approved") or state["revisions"] >= 2:10        return Command(goto="__end__")11    return Command(goto="critic")1213builder.add_node("supervisor", supervisor)14builder.add_node("researcher", research_agent)   # create_agent with search tools15builder.add_node("writer", writer_node)          # strong model, no tools16builder.add_node("critic", critic_node)          # cheap model, rubric, structured output17for role in ("researcher", "writer", "critic"):18    builder.add_edge(role, "supervisor")19builder.add_edge(START, "supervisor")

Supervisor as an LLM with subagents as tools

1from langchain.agents import create_agent2from langchain.tools import tool34researcher = create_agent("openai:gpt-4.1-mini", tools=[web_search],5                          system_prompt="Find facts with sources. Return bullet points.")67@tool8def research(question: str) -> str:9    """Research one question and return findings with sources."""10    out = researcher.invoke({"messages": [{"role": "user", "content": question}]})11    return out["messages"][-1].content1213supervisor = create_agent("anthropic:claude-sonnet-4-5", tools=[research, write_section])

The supervisor sees only the researcher's final answer, not its 12 search steps, which keeps its context small.

Design rules

• Narrow shared state — plan, findings, draft, critique, revisions.
• Tools per role — a writer with a search tool will search instead of writing.
• Structured critic — {"approved": bool, "issues": [...]}.
• Deterministic routing where possible — an LLM supervisor runs on every hop and is often the most expensive part.

A real-life example

A consulting firm's "sector brief" system has four roles. The planner turns "EV two-wheeler market in India" into five questions; the researcher (a create_agent with search and a filings tool) answers each; the writer drafts 1,200 words from the findings only; the critic checks every number has a source. The supervisor is 12 lines of code. An early version used an LLM supervisor, which sometimes sent the draft back to research for no reason; switching to code cut average cost from Rs 42 to Rs 27 per brief with no quality change.

Follow-up questions to expect

• "Code supervisor or LLM supervisor?" — Code when the order of roles follows clear rules; an LLM when the next step depends on understanding the request.
• "Should each role be a subgraph?" — Only roles with their own loop or tools. A single model call is just a node.
• "How do roles share memory?" — Through typed state keys; each role reads only what it needs, formatted into its own prompt.