Multi-agent system: several AIs sharing the work
A multi-agent system has several AI agents work together, each with its own role, instructions and sometimes tools: one coordinates, others search, write or check. This setup helps with big tasks that can be split up, but it uses far more resources.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
The idea: a small team instead of one generalist
A single AI agent has to keep everything in its context window, the amount of text it can read at once. In a multi-agent system, the task is split: a lead agent drafts a plan, hands subtasks to specialized agents, then pulls their results together. Getting software agents to cooperate is an old idea in computer science; it gained new momentum with large language models, notably with Microsoft’s AutoGen framework, introduced in an August 2023 paper.
A concrete example
A nonprofit wants a summary of every public grant available for home energy upgrades.
- The coordinator splits the question: federal programs, state programs, income requirements.
- Three research agents each explore one track in parallel, running their own web searches.
- A checker agent compares the findings with official pages.
- The coordinator assembles a single summary.
Each agent works with a shorter, more focused context, and the searches run at the same time.
Why it matters, and what it costs
In June 2025, Anthropic described its own multi-agent research system: a lead agent that spins up subagents in parallel. The company notes that this kind of system uses about 15 times more tokens than a chat, and that it suits high-value tasks that parallelize well or exceed what a single context can hold. To see how that multiplier affects a bill, try our agent cost calculator.
Common mistakes and myths
- “More agents means more intelligence.” More agents mostly means more messages, more cost and more room for misunderstandings between them.
- “Agents understand each other perfectly.” Instructions passed from one agent to another can be vague or incomplete; this is a frequent source of errors.
- “Multi-agent and orchestration are the same thing.” Orchestration is how the agents are coordinated; the multi-agent system is the whole setup.
Related terms: agentic workflow, A2A protocol, task decomposition.
Related terms
- AI agent · An AI that does more than answer: it plans and carries out a series of actions with tools to reach a goal you set.
- Agent orchestration · Coordinating several model calls, tools or AI agents: who does what, in which order, and how the results are brought together.
- Agentic workflow · A sequence of steps in which an AI model plans, uses tools, checks and revises its work instead of answering in a single pass.
- Agent2Agent protocol (A2A) · An open protocol that lets AI agents from different vendors discover each other, communicate and hand tasks to one another.
- Task decomposition · Breaking a complex goal into smaller subtasks that a model or AI agent can handle one at a time or in parallel.
- Context window · The maximum amount of text (in tokens) a model can read and produce at once, including its own answer.
Try it
Full glossary
Frequently asked questions
Does a multi-agent system use several different AI models?
Not necessarily. Often it is the same model called several times with different instructions and tools. You can also mix models, for example a powerful one to coordinate and a lighter one for subtasks.
Why does a multi-agent system cost more?
Each agent reads its own instructions, documents and messages from the others, which multiplies the tokens used. Anthropic reports that its multi-agent research system uses about 15 times more tokens than a regular chat.
Which tools can I use to build one?
Open-source frameworks such as AutoGen, CrewAI or LangGraph are designed for it. Some no-code platforms also offer agents that can call each other.
Is it always better than a single agent?
No. When every step depends closely on the previous one, a single agent is often simpler, cheaper and just as effective.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.