Building a Multi-Agent Customer Service System with LangGraph — A Practical Guide

This is not a toy example. We are going to build a multi-agent customer service system where different AI agents handle different types of enquiries — greetings, product queries, pricing, bookings, policy questions, and escalation to humans. By the end, you will have a working LangGraph state machine that classifies intent, routes to the right agent, retrieves relevant context via RAG, and generates a response. The code here is drawn from Tork Chat , a production multi-agent assistant deployed in the vehicle rental industry. I have simplified some of the production concerns (governance, multi-tenancy, observability) to focus on the multi-agent pattern itself. Why multi-agent over single-prompt A single prompt can answer a single question well. It falls apart when a customer does this: "What SUVs do you have?" → fleet query "How much for 3 days?" → pricing "OK book it for next Friday" → booking intent "What's your fuel policy?" → policy lookup "Actually this is too expensive, let me spe