Multi-Agent System

In computer science, an agent is a program that pursues a task independently: it observes its environment, decides on its own what the next step should be, and acts. A multi-agent system consists of several such programs that run at the same time and have to interact with one another. They exchange messages, divide up work, or compete for the same resources. What matters is this: there is no single place that dictates everything down to the last detail. Each agent makes its own decisions, and the behavior of the whole only emerges from the interplay between them. You can picture a kitchen with several cooks who are supposed to deliver the same menu without every single step being planned out in advance.

Why tasks are split across multiple agents

Large tasks almost always break down into subtasks. A single program meant to handle everything at once quickly becomes unwieldy and error-prone. If you distribute the parts across specialized agents, each part stays manageable. If one agent fails, the others can, in the best case, keep working.

In the current AI debate there is a second reason. Language models—that is, programs like ChatGPT that continue text—make mistakes over long chains of tasks. They lose track or invent details. If a second model checks the work of the first, some of these errors come to light. This division into a worker and a checker is the most common reason companies deploy several agents instead of just one.

However, the benefit shouldn’t be overestimated. Every additional agent costs computing time, and thus money. Five agents sending each other intermediate results can easily consume ten times what a single pass would. And if all the agents are based on the same model, they often share the same blind spots.

Roles, messages, and a coordinator

Technically, each agent is assigned a role and a goal. For AI agents, this happens via an instruction in plain language, such as: “You are the research agent, you gather sources and pass them on.” In addition, each agent is given tools it is allowed to access. Typical tools include a web search, a calculator, or a database.

The agents talk to each other via messages. One writes its result to a shared memory, the next reads it there and continues working. There is often an additional coordinator, frequently called an orchestrator. It breaks down the user’s request, distributes the parts, and assembles the answers at the end.

The tricky question is when to stop. Without a termination rule, two agents can correct each other endlessly. That’s why the number of rounds is limited, or a check condition is defined, upon whose fulfillment the system delivers its output. A related term is the agent pipeline: there, the order of steps is fixed in advance. A genuine multi-agent system, by contrast, decides for itself who goes next.

From delivery robots to coding assistants

The term is older than the current AI boom. In logistics centers, transport robots steer around each other without a central control planning every trip. In the power grid, programs negotiate supply and demand among many small producers. High-frequency trading on stock exchanges is also a multi-agent system, in which programs from different banks react to one another.

In products today, the term is encountered mainly in coding assistants and research tools. A coding assistant might have one agent write code, a second run tests, and a third evaluate the error messages. “Deep research” tools send several agents out to different sources in parallel and summarize the findings.

In business news, the term usually appears as a promise: entire workflows are supposed to be taken over by teams of agents in the future. When reading such claims, it’s worth asking a critical question: does anyone check the results before they take effect? As long as agents make mistakes, this oversight is needed—especially when money is being transferred or contracts are being concluded.