Module 5 · Project 4: Multi-Agent Customer Support System

OBJECTIVE: build a customer-support system where a supervisor agent routes each customer message to specialist agents — orders, billing and policy/help — which use the tools from earlier projects, hand off to each other, and escalate to a human when needed. Learn how to design, build, test and control multi-agent systems. WHAT WE COVER (in project order): • When multiple agents help (and when they hurt) • Multi-agent patterns: supervisor, hierarchical, handoffs, group chat, debate • Designing roles, tools and handoff rules • Building with LangGraph (supervisor + specialists) and a CrewAI version for comparison • Shared state, coordination and simple consensus (voting) • Agent-to-Agent protocol (A2A) — connecting agents across systems • Evaluation, tracing, failure modes, and cost/latency optimisation • Security and guardrails for a multi-agent system • Deployment as a chat app AFTER THIS MODULE YOU CAN: • Design a multi-agent system with clear roles and hand-offs • Build it in LangGraph and CrewAI and choose between them • Test, trace and control cost of a multi-agent system DATA: the shop database (via MCP), policy documents (agentic RAG from Project 3) and 60 test conversations you write. TOOLS: LangGraph, langgraph-supervisor, CrewAI, A2A SDK, LangSmith/Langfuse, Streamlit.