What is Multi-Agent System?
Also called MAS, agent team.
A multi-agent system is one in which several AI agents, each with its own instructions, tools, and scope, work on parts of a larger goal and exchange information. Agents may run in sequence, in parallel, or under a coordinator. The design trades the simplicity of one agent for specialization, isolation of context, and parallel progress.
The usual reasons to split are context and focus. A single agent given twenty tools and a long brief tends to lose the thread, pick the wrong tool, or fill its context with irrelevant history. Separate agents each carry a narrower instruction set and a shorter working context, which measurably improves reliability on tasks with distinct phases.
Several arrangements are common. A supervisor pattern has one agent that decomposes work and delegates to specialists. A sequential pipeline passes output down a chain. A parallel pattern fans identical work across agents and merges results. A peer network lets agents message each other freely, which is the most flexible and by far the hardest to debug.
Coordination is not free. Every handoff loses information that was not written down, agents can duplicate work or deadlock waiting on each other, and total token cost rises with the number of participants. Debugging is harder because a wrong final answer may originate several agents upstream. Many teams that start multi-agent consolidate back to fewer, better-scoped agents.
Multi-agent systems are often confused with the separate idea of one agent that has many tools. The distinction is whether each component holds its own instructions and reasoning context. Related concepts include orchestration, which is the machinery that routes work, delegation and handoff, which describe transfers between agents, and agent trajectory, which is how a run is inspected afterward.
Key points
- Splitting work keeps each agent's context and tool set narrow.
- Common shapes: supervisor, sequential pipeline, parallel fan-out, peer network.
- Handoffs lose whatever context was not explicitly written down.
- Coordination adds cost, latency, and harder debugging.
In practice
A market research request is split across three agents. One gathers recent news, one pulls pricing pages, and one reads customer reviews. Each returns a short structured summary. A fourth agent reads all three summaries and writes the final brief. None of them sees the others' raw material, only the summaries, which keeps each working context small and focused.