LangGraph Multi-Agent

When to use LangGraph

• Long-running sessions need checkpointing and resume
• Branching logic (if/else) must be explicit, not prompt-only
• Human-in-the-loop (HITL) approvals between steps

Minimal graph pattern

from langgraph.graph import StateGraph, START, END
from typing import TypedDict, Annotated
import operator

class State(TypedDict):
    messages: Annotated[list, operator.add]

def planner(state: State) -> State:
    # call LLM; append message
    return state

def tool_executor(state: State) -> State:
    # run tools; append results
    return state

builder = StateGraph(State)
builder.add_node("planner", planner)
builder.add_node("tools", tool_executor)
builder.add_edge(START, "planner")
builder.add_conditional_edges("planner", route_fn, ["tools", END])
builder.add_edge("tools", "planner")
graph = builder.compile(checkpointer=checkpointer)

Workflow

- [ ] Define State TypedDict (only fields you need)
- [ ] One node per responsibility (plan, act, reflect)
- [ ] conditional_edges for routing (not mega-prompts)
- [ ] compile(checkpointer=...) for persistence
- [ ] interrupt_before for HITL on sensitive nodes

MCP integration

• Tools live on MCP servers; graph nodes call MCP client, not ad-hoc HTTP
• Keep orchestration in LangGraph; keep side effects in MCP tools
• Pair with multi-agent-supervisor when multiple role-specific subgraphs exist

Observability

• Tag runs with thread_id and user_id (hashed)
• Export traces to LangSmith or OpenTelemetry (see agent-observability)

Common mistakes

• Storing full chat history in state without trimming → use summarization node
• Cyclic graphs without max step guard → set recursion_limit
• Single node doing plan + act + evaluate → split nodes