A multi-agent system uses two or more agents with defined roles or responsibilities. One agent might research, another draft, and a third review the result. The agents may share a task board, exchange messages or pass structured outputs between steps.
A common example is a content pipeline: a research agent collects sources, a writer agent drafts from those notes, and a checker agent compares the draft with the sources and flags unsupported claims. Each agent gets a narrow instruction and a limited set of tools, which makes its behavior easier to test than one large prompt that does everything. A shared task list or agent state keeps the handoffs visible.
More agents do not automatically mean better results. Coordination adds latency, token cost and new failure modes, so begin with one agent and split responsibilities only when testing shows a clear benefit. See AI Agent and AI Agent Memory.