Multi-agent orchestration with LangGraph

Single-prompt agents fall over once a task needs distinct skills. LangGraph models the workflow as a typed state graph: a router decides which specialist runs next, each node owns one job.

graph TD;
  Q[Query] --> R{Router};
  R -->|retrieve| V[Vector search];
  R -->|answer| A[Answer];
  V --> A;
  A --> D([Done]);

LLM-as-router dispatching to specialist nodes

graph.py

from langgraph.graph import StateGraph, END

graph = StateGraph(AgentState)
graph.add_node("router", route)
graph.add_node("retrieve", retrieve)
graph.add_node("answer", answer)
graph.add_conditional_edges("router", pick_next, {"retrieve": "retrieve", "answer": "answer", "done": END})
app = graph.compile()

Where it breaks

Shared mutable state across agents is the usual culprit. Keep node outputs additive and validate the state shape at each edge.

RAG ingestion to retrieval architecture — Ingestion → pgvector → retrieval