LangGraph @ Work: Building Production AI Agents

Your LangGraph agent runs perfectly on your laptop. Then you deploy it. The conversation memory vanishes on restart. Refunds approve themselves with no human in sight. And when it breaks in front of a customer, you have no way to see why.Most LangGraph books either stop at the RAG demo or sprawl across the entire LangChain ecosystem. This one does neither. LangGraph @ Work is the production half that almost nobody teaches - and it gives you the code.This is a hands-on, build-it guide that takes you from the agent mental model to a fully operational multi-agent system in Python. One real-world example, developed end-to-end, every line in your hands by the finish. What you'll build. A complete customer service agent. A supervisor routes requests to specialized catalog and billing subagents. A refund flow pauses for human approval via interrupt(). PostgreSQL-backed memory persists across sessions. A FastAPI gateway handles concurrent async requests. You run it. You ship it.What you'll learn: - LangGraph state, nodes, and edges built from primitives - not from black-box helpers- The supervisor pattern with @tool-wrapped subagents that keep context windows small as the system grows- Where to place interrupt() for human-in-the-loop, and why node-versus-tool is a real architectural decision- The InMemoryStore-to-PostgresStore swap, with the connection pooling and setup() idempotency production demands- Full observability with LangSmith: trace every decision, run evaluations in CI, and build an LLM-as-judge with a working rubric- Testing graphs with mocked LLM calls and in-memory SQLite fixtures- Async FastAPI deployment with ainvoke(), lifespan management, and CORS allowlists- Dependency security and the recent LangChain CVEs, with the exact versions to pin in requirements.txtFree source code. Every line of the worked example is a free download - modules, database schema, setup scripts, sample data, and the complete security-pinned requirements.txt.Written for the individual engineer who has to ship it, not the team still discussing it. Software engineers, ML engineers, AI engineers, solutions architects, and the rising class of Forward-Deployed Engineers building real AI products, not demos.About the author. John Moore is a learning designer and technical author with two decades of experience building engineering curricula for major cloud platforms, including the most recent Architecting on AWS update. He writes for developers who want to move fast and build things that work.Stop reading about agents. Start building them.