A multi-agent system splits work across several agents, each with a narrower role, tool set, or context — a planner, a researcher, a writer — that hand off to each other. The pitch is specialization: a focused agent with the right tools beats a generalist juggling everything.
Why it matters in production
Decomposition can genuinely help. A smaller context per agent means less for the model to lose track of, and role separation makes each step easier to evaluate. But every hand-off is a new interface where information gets dropped, and every extra agent multiplies latency and cost. The coordination overhead is real, and it is the thing people underestimate.
The common mistake
Adding agents to feel sophisticated. Two agents that pass a vague summary between them are often worse than one agent with a clear plan, because the summary lost the detail the next step needed. Multi-agent is a tool for genuine parallelism or genuinely distinct skill sets — not a default architecture.
How we actually use it
We reach for multiple agents only when the sub-tasks are truly independent or need incompatible tool sets, and we make the hand-off contract explicit rather than hoping a free-text summary carries enough. We wrote up the real structure of one such build in the anatomy of a multi-agent platform, and the context engineering that keeps each agent's window clean is what makes the whole thing hold together.