multi-agent systems
/ MUL-tee-AY-junt SIS-tum /
A multi-agent system is several AI agents working together on one problem, often each playing a different role — like a small office where one person researches, another writes, a third checks the work, and a manager hands out the tasks and stitches the results together. Instead of one model trying to do everything, the job is divided among collaborators that talk to each other (usually by passing messages of text).
The appeal is divide-and-conquer plus specialization. Different agents can be given different tools, instructions, or even different underlying models, and having one agent critique or verify another's output can catch mistakes a single pass would miss. The pattern borrows ideas from an older research field of the same name, where independent software "agents" negotiate and coordinate, and from how human teams split labor.
But the costs are easy to underestimate, and the field is candid about this. Every agent is another model running, so multi-agent setups can be several times slower and more expensive than a single call. Errors don't just stay put — a wrong fact from one agent can be confidently passed along and amplified by the next. Agents can also misunderstand each other, loop in circles, or all agree on the same mistake. In practice, a single well-designed agent often beats a crowd, and you reach for multiple agents mainly when a task genuinely splits into independent parts.
To write a market report, a "researcher" agent gathers data with a search tool, a "writer" agent turns it into prose, and a "reviewer" agent checks the prose against the data and flags unsupported claims. A coordinator routes the work and assembles the final document. If the reviewer catches a fabricated figure, it sends the section back to the writer — a check a lone agent would likely have skipped.
Specialized roles that check each other — at the cost of more model calls.
More agents is not more intelligence. Several agents share the same weaknesses if they share the same model, so they can confidently reinforce one another's errors instead of correcting them. The real benefit comes from genuine specialization and independent checking — not from simply adding more voices to the room.