Multi-Agent System

A multi-agent system consists of several programs that solve a task together. Each of these programs is called an agent. An agent is given a goal, decides on its own next steps, and may use tools for this, such as a search engine or a calculator. In a multi-agent system, such an agent does not work alone but passes results on to other agents. You can picture it like a project team: one researches, one writes, one checks at the end. This division of labor is not an end in itself but is meant to deliver better results than a single all-rounder.

Why tasks are distributed across multiple agents

Large tasks often fail because too much has to be considered at once. A language model that is supposed to research, calculate, formulate, and check all at the same time easily loses the thread. If you break the task down, each agent gets a narrow, clearly defined role. Narrow roles are considerably easier for today’s models to fulfill reliably.

A second advantage is mutual oversight. AI models occasionally invent facts that sound convincing but are wrong. A dedicated checking agent that only looks for contradictions and missing evidence finds such errors more often than the model does on its own. Specialist companies, too, therefore rely on separate roles rather than a single large command.

The price for this is effort. Each agent incurs its own computing costs, because the model runs anew for each step. A multi-agent system can easily cost five to ten times as much as a single call for the same question. That’s why the approach is worthwhile mainly for complex tasks, not for a quick translation.

Roles, messages, and the orchestrator

In most systems there is an overarching instance, the orchestrator. It receives the task, breaks it down into subtasks, and distributes them. The sub-agents work through their assignments and send results back. The orchestrator assembles the answer from these or starts another round.

Communication mostly runs via text. One agent writes down its result, the next reads it as part of its task description. Technically, the same language model often sits behind all agents. The only difference is the instruction with which it is started and the tools it is allowed to use.

A typical misconception is the idea that the agents would discuss with each other like humans. In fact, they follow a fixed procedure specified by developers. Where this procedure is too loose, agents go around in circles or reinforce an error together. That’s why developers build in stopping rules, such as a maximum number of rounds.

Where such agent teams are in use today

Multi-agent systems are most visible in programming. Tools like Devin or the agent modes in Cursor and GitHub Copilot plan a change, write code, run tests, and make corrections. Each of these steps can be taken on by its own agent. The research features offered under names like Deep Research work in a similar way.

In companies, such systems appear in customer support and in accounting. One agent reads incoming invoices, a second matches them against orders, a third reports discrepancies to a human. This last step remains important: when it comes to money and contracts, a human almost always gives the final approval.

In the news, you often encounter the term in connection with the word agentic. This means that AI not only responds but acts on its own. The multi-agent system is to be distinguished from the simple chatbot, which only reacts to individual questions, and from the single agent, which does act but works without colleagues.