Build your first production agent with LangGraph

This guide ships a small but real agent. We will build a research assistant that can browse, summarize, and persist findings — with retries, idempotency, and traces.

What you will build

A single-graph LangGraph agent with three tools: search, fetch, and append-note. State is persisted with a SQLite checkpointer so runs survive crashes.

Step 1: Project scaffold

Use uv or poetry. The dependency footprint is intentionally small.

Step 2: Define your tools

Wrap each tool with input/output validation. Most production failures originate at this layer.

Step 3: Wire the graph

Keep the graph shallow — three nodes, one router. Avoid the temptation to add a critic on day one.

Step 4: Add tracing

LangSmith or Langfuse, your choice. The point is to see every tool call.

Step 5: Run replay tests

Capture five real sessions. Replay them on every change. This is the cheapest insurance you will ever buy.