What are multi-agent systems?
One agent has limits. A multi-agent system is several specialized agents — each with a focused role and its own tools — collaborating on a workflow too big for any one of them. Learn how they're wired, where they break, and how to keep them under control. Builds on the AI Agents lesson.
01How they're wired: three topologies
One skilled generalist can handle a small job alone, but a big project usually goes faster with a team of specialists who each own one part and hand work to each other. Software works the same way: a single AI worker can chase one goal by itself, while a multi-agent system splits a large job across several AI workers — each with a focused role and its own tools — then wires them together. The wiring — the topology — is the big design choice. Flip between the three most common shapes.
A lead agent delegates sub-tasks to specialist workers, then combines their results into one answer.
- Specialization — each agent gets a narrow role and just the tools it needs, so it does that one job well.
- Parallelism — independent sub-tasks can run at the same time instead of one long chain.
- Modularity — swap, add, or upgrade a single agent without rewriting the whole system.
02The roles on a team
In a supervisor–worker team, each agent is a specialist. Tap each role to see what it owns — notice the supervisor never does the work itself; it plans, delegates, and assembles.
Supervisor
The lead agent. It breaks the goal into sub-tasks, hands each to the right specialist, then combines their results into the final answer. It coordinates rather than doing the hands-on work itself — which makes it the natural place to put oversight and limits.
03A multi-agent task, step by step
Watch a supervisor–worker team handle one goal end to end: "Write a short brief on a new technology." Step through it, or run the whole thing.
04The trade-off: power vs. new failure modes
Splitting work across agents buys real benefits — but coordination also creates failure modes a single agent never had. Switch between them.
Benefits — why split the work at all
Specialization, parallelism, and modularity. Each agent has a narrow role with just the tools it needs, independent sub-tasks run side by side, and you can upgrade one agent without touching the rest.
Cascading failure — one error spreads
Because agents feed each other, a single mistake or hallucination upstream can be trusted and amplified downstream. If the researcher returns a wrong fact, the writer builds on it and the critic may miss it — the error propagates through the whole chain.
Runaway loops — endless sub-tasks
Agents that can spawn more agents or re-delegate can loop without converging — endlessly creating sub-tasks, re-trying, or handing work back and forth. Left unbounded, this burns time and budget with no result.
Coordination cost — complexity adds up
More agents means more messages, hand-offs, and points of failure. Coordinating who does what, passing context cleanly, and debugging across agents is harder than running one loop. The system can be slower or pricier than a single well-scoped agent for simple tasks.
05Wizard: is your multi-agent system safe to run?
The same controls from the AI Agents lesson scale up here — and they matter more, because a multi-agent system has more ways to spiral. Answer five questions about your setup and get an instant read with the specific controls to close any gaps.