AI glossary · Basic · Agents
Multi-agent
Multi-agent is an approach in which several AI agents split up roles and exchange results with one another to handle one large task together. Instead of one agent doing everything alone, they work like a team.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents
- Added
- 2026-10-08
How it works
A common setup is for a coordinating agent (an orchestrator) to break a user's request into small tasks and assign them to other agents. Each agent has prompts and tools suited to its role; for example, one handles web search, one writes code and one reviews results. The results each one produces are gathered back at the orchestrator, which combines them into the final answer. There are also setups without an orchestrator, where agents pass work to each other in turn or exchange views as if in a debate.
Why it matters
Giving a single agent long instructions and many tools all at once fills up the context window quickly and increases mistakes. Splitting roles means each agent only needs to see the information required for its own task, which improves focus. Independent tasks can be handled by several agents at the same time, which saves time. Having one agent review another agent's results also helps catch mistakes.
Good to know
As the number of agents grows, the number of model calls grows too, increasing cost and time. Errors can pile up as one agent's wrong result is passed to the next. Designing what is exchanged between agents, and in what format, has a large effect on performance. So the common recommendation is to handle simple tasks with a single agent and use multi-agent setups only for complex tasks that are truly worth splitting. Standards such as A2A, which aim to let agents built by different companies communicate with each other, are also emerging.
Example
For a service that produces research reports, it could be set up like this: an orchestrator splits a topic into several sub-questions, several research agents each find material, a writing agent organizes it into a report, and a review agent checks the facts. As in the news item on this site, 'Docker releases Docker Agent, a CLI plugin for defining and running AI agents with YAML,' tools that define and run agents' roles and tools from configuration files are making such setups easier to build and manage.