Single-agent systems hit practical limits. A long task may require research, planning, implementation, review, execution, and synthesis. Putting all of that into one prompt and one context window often produces bloated state, weak verification, and hard-to-debug behavior.

A multi-agent design splits the work across specialized agents with explicit handoffs.

A multi-agent system consists of several agents that coordinate to solve a larger task. Each agent owns a role, toolset, context, or phase of the workflow. A researcher may gather sources, a writer may synthesize, a verifier may check claims, and a supervisor may decide when the result is ready.

This does not automatically make the system more reliable. Multi-agent systems add coordination cost, latency, and new failure modes. They are useful when specialization, parallelism, independent review, or context isolation is worth that overhead.

This chapter covers the patterns that matter in practice: supervisor-worker, peer-to-peer loops, debate and critique, structured handoffs, and the decision criteria for when multi-agent design beats a simpler single-agent system.