What is Multi-agent system?

What it is

A multi-agent system uses more than one AI agent on the same goal. Each agent runs its own loop of planning, acting and checking, and the agents pass information to one another. The agents may be copies of the same model with different instructions, or different models chosen for different jobs.

The nearest everyday picture is a small team. One person researches, one writes, one reviews, and they talk to each other along the way.

How the agents are organized

There are a few common layouts:

• Orchestrator and workers. A lead agent breaks the task into pieces, hands them to worker agents, and combines the results.
• Specialists. Each agent has a role, such as planner, coder or tester, with its own instructions and tools.
• Peers. Agents of equal standing coordinate directly, for example through a shared chat channel or shared notes.
• Reviewer pairs. One agent produces work and another critiques it before it counts as done.

Whatever the layout, agents need a way to share state. That is often a message channel, a shared memory, or a common file area. Each agent has its own context window, so splitting work can keep any one of them from filling up with irrelevant detail.

Why people build them

Splitting a job helps in three ways. Work that is independent can run in parallel, which shortens the clock time. Narrow roles make each agent's instructions simpler. And a second agent checking the first can catch mistakes that one agent would repeat.

The approach suits long tasks such as large agentic coding projects or research that touches many sources.

The costs

More agents is not automatically better. Every message between agents costs tokens and time, so the bill grows with the team. Coordination can fail in ways one agent cannot: two agents edit the same file, one trusts a wrong claim from another, or they loop politely at each other. A mistake or injected instruction in one agent's output can spread to the rest, so the risks of prompt injection multiply.

For many jobs a single well-instructed agent is cheaper and easier to debug. A sensible rule is to add agents only when you can name what the extra one does that the first cannot, and to measure the result with evals against the simpler setup. Keep logs of who did what, and run the group inside an agent sandbox with limits you control.