Multi-agent system

A multi-agent system is software made up of several independent sub-programs. Such sub-programs are called agents: they are given a goal, make their own decisions, and can use tools, such as a search engine or a calculator. What matters is that no one dictates every single step. You tell an agent what the outcome should be, not how to get there. In a multi-agent system, several such agents work on the same task. In doing so, they exchange intermediate results, split up the work, and partly check each other. The comparison to a project team fits well: multiple participants, clear roles, one shared outcome.

What a team of agents can do better than a single one

A single agent quickly hits its limits on long tasks. It loses track, repeats steps, or gets stuck on an error. If the task is split up, each agent only handles a manageable portion. This noticeably lowers the error rate, because each sub-task is formulated more clearly.

A second advantage is mutual control. Language models occasionally invent facts that sound plausible but are wrong. If a second agent is deployed as a checker, such an error is caught more often. This is no guaranteed safety net, but it helps measurably.

Then there is speed. Several agents can work simultaneously, for example evaluating ten sources in parallel. A single agent would have to go through them one after another. For research tasks, this saves a lot of time. The price for this is high, however: each agent incurs its own compute costs, and a team of five agents can easily cost five times as much.

Roles, messages, and a coordinator

Most of the time, everything starts with a coordinating agent, often called an orchestrator or planner. It breaks the big task down into sub-tasks and distributes them. The other agents are assigned fixed roles, for example researcher, programmer, or reviewer. This role is embedded in an instruction that is permanently given to the agent.

The agents communicate via messages, technically usually simple blocks of text. One agent writes down its result, another reads it and continues working with it. Often there is also a shared memory for intermediate states, so that not every piece of information has to travel through all the messages. In the end, the coordinator compiles the sub-results into a final result.

A typical misconception is the idea that a separate AI model lies behind each agent. Usually it is the same model, just with different role instructions and tools. The difference lies in the task, not in the technology. And these systems are prone to failure: if one agent misunderstands its assignment, the error propagates through the entire chain.

From research tools to the supply chain

Agent teams are currently best known in programming. Tools like Devin or the agent features in common programming environments let one agent write code, a second test it, and a third fix bugs. The deep-research features of major chatbots work similarly: several agents search the web in parallel and ultimately deliver a summarized report.

Outside the world of AI, the idea is much older. In logistics, software agents have been negotiating delivery times and storage space for years. In the power grid, agents balance generation and consumption across many small facilities. Traffic simulations, too, compute each car as its own agent.

In business news, the term mainly appears as a sales pitch. Companies promise to hand over entire workflows to agent teams. A sober look is worthwhile: many of these systems work well on clearly defined tasks and fail on open-ended ones. Anyone reading such reports should therefore ask what task was actually solved and who ultimately checks the result.