What is a multi-agent system?

Picture a restaurant kitchen at the dinner rush. One person calls the orders. Someone works the grill, someone makes the salads, and someone plates each dish before it goes out. Nobody could cook the whole menu alone at that speed. Some artificial intelligence (AI) tools now split a big job the same way, among several agents. That setup is called a multi-agent system.

An agent is AI software that works toward a goal in steps, using tools such as search and checking what happens as it goes. Topic 32 covers the idea. A single agent does everything itself, one step after another, like one cook working alone. A multi-agent system is several agents, each with its own job, working on one larger task. Often one of them coordinates the rest.¹³

Say you ask a research tool to compare how three countries label food. In a multi-agent setup, a lead agent reads your question first and makes a plan. It breaks the job into parts, one country each, and writes instructions for every part. Anthropic, the company that makes the Claude models, describes its own research system this way: a lead agent coordinates the job and hands pieces of it to helper agents.²

Each helper takes its part and gets to work, often at the same time as the others. One searches for the first country's rules while another reads about the second. Working side by side is the main appeal, because three searches at once finish sooner than three in a row. And each helper can focus on its own piece without carrying everything the others found.²

When the helpers report back, the lead agent reads what they found, notices gaps, and may send a helper out again. Then it combines the pieces into one answer. Like the person who plates each dish, it's the last stop before anything reaches you. If you've used a deep research mode that searches many pages and writes a report, you may have met this already, because some of those tools work this way behind the scenes. Topic 63 covers deep research.²

All that teamwork costs something. Every agent reads, thinks, and writes, and that computing is measured in tokens, the small pieces of text a model works in. Anthropic reported in 2025 that its multi-agent system used about fifteen times as many tokens as a chat. So builders tend to save teams of agents for jobs worth the expense, like broad research. Some jobs split badly, too. When every step depends on the one before, extra helpers mostly wait.²

Each hand-off between agents is another place for a mistake to slip in. If a helper misreads its instructions, or brings back a wrong fact, the lead may build on it, and the error can travel all the way into the final report. Anthropic's write-up says errors in agent systems compound: one failed step can send the whole job down a different path. So when a tool says a team of agents did the work, open two or three of the sources it lists. Do they say what the report says they do?²