Agent Swarm
An agent swarm is a group of AI programs that work together on a task instead of a single program doing everything on its own. Each program takes on a sub-step, passes on its result, and sometimes checks the work of the others.
Some computer programs today can independently carry out several work steps in sequence. They are given a goal, work out intermediate steps for themselves, and use tools such as a search engine or a calculator to do so. Such programs are called agents. An agent swarm, then, is not a single piece of software of this kind, but an entire group of them working together on a task. The members split the work among themselves and send each other their intermediate results. The comparison with a project team fits well: one plans, one researches, one checks at the end whether the result is correct.
What a team can achieve that an individual cannot
A single agent easily loses track on long tasks. If it has to perform twenty steps in a row, something already goes wrong by step three. This error then travels through all the following steps. In the end the result is wrong, without anyone noticing where the error arose.
A swarm can cushion this weakness. One agent produces a draft, a second reads it critically and reports contradictions. This division of roles works surprisingly well in practice. The checker has only one task and is not distracted by the fact that it wrote the text itself.
On top of that, there is speed. Ten agents can read ten sources at the same time instead of working through them one after another. This is exactly why the term is so prominent in the business press. Companies hope that this will make it possible to automate entire workflows that previously required human coordination.
Roles, messages, and a lead agent
Usually the same language model underlies all the agents. This is a system that has learned to continue texts, and can thereby answer questions. What differs is only the instruction each agent receives at the start. One is told: You are a researcher and look for evidence. Another: You are an editor and look for errors. This initial instruction defines the role.
There are two common designs for collaboration. In the first, there is a superordinate agent, the orchestrator. It breaks the goal down into sub-tasks, distributes them, and combines the answers at the end. In the second, the agents talk directly to one another, without a boss. This is more flexible, but tends to get out of control more quickly.
The exchange happens via text. One agent writes its result, the next one receives it as input. This is exactly where the biggest problem lies: every message costs computing time and thus money. Twenty agents all talking to each other produce a very large number of messages. That is why developers limit the number of rounds after which the swarm must break off.
From coding assistants to market analysis
Swarms are most visible in software development. Tools like GitHub Copilot or Claude Code let several agents work on a project in parallel. One writes program code, one writes tests for it, one brings the documentation up to date. Something comparable exists for research: one agent collects studies, others summarize them, one writes the report.
You encounter the term in financial news when banks or consultancies talk about automation. What is meant, for example, are processes in which agents evaluate reports, compare figures, and produce a summary. Such announcements often sound more advanced than the technology actually is.
A common misconception: more agents do not automatically mean better results. If they all rely on the same model, they also share its weaknesses. If the model invents a source, the checking agent may likewise take it to be real. A swarm therefore does not replace human oversight at the decisive points.