Agents working as a team
A multi-agent system has several AI agents split the roles and hand information and results to each other in pursuit of a single goal.
One agent researches, another writes, another checks for errors. Instead of one model doing everything, agents with different strengths and permissions are combined like a team.
Ways to divide the work
- A planner decomposes the job and assigns owners
- Several researchers work different sources in parallel
- A producer and a reviewer pass work between them
- Text, images, and code go to specialized models
- Consequential actions return to a human approver
Some arrangements run agents in sequence; others run many at once. Features that fan out large numbers of agents to research or process in parallel now exist as well.
Compared with one agent, or a workflow
A single agent plans, executes, and checks by itself. That is simpler and keeps cost and latency down, so trying one agent first is usually the rational move.
A workflow fixes the order and the conditions in advance. In a multi-agent system, the AI may decide who acts next. In practice the two are combined: a workflow provides the safe path, with several agents operating inside it.
More agents is not automatically better
Each additional agent re-reads the same context, increasing tokens, cost, and latency. Information gets dropped in handoffs, and different agents can reach contradictory conclusions.
Define the roles, what gets passed, the stopping condition, and where failures return to — and keep logs you can inspect. One agent for simple work; multiple agents where permissions and specialisms genuinely need separating.