What is a multi-agent system?

A multi-agent system uses multiple AI agents working together to accomplish complex tasks. Each agent specializes in a specific capability — one researches, one writes, one reviews, one executes actions — and they communicate to complete workflows that no single agent could handle well alone. Architecture patterns: 1) Sequential pipeline — Agent A processes input, passes to Agent B, which passes to Agent C. Simple, predictable. 2) Supervisor pattern — a coordinator agent assigns tasks to specialized workers and aggregates results. Good for parallelizable work. 3) Debate/critique — two agents review each other output, improving quality through adversarial collaboration. 4) Swarm — multiple equal agents collaborate with shared state, self-organizing around subtasks. Examples: a research agent that gathers information + a writing agent that produces content + a review agent that checks quality. A customer service system where a triage agent routes to specialist agents (billing, technical, account management). Current state: multi-agent systems work well for structured workflows (research → write → review) but struggle with open-ended coordination. They are more expensive (multiple LLM calls per task) and harder to debug (agent interactions create emergent behavior). Start with single-agent systems. Add agents only when a single agent demonstrably fails at your task complexity.