What is a multi-agent system in AI?

Also called MAS, multi-agent architecture

Definition

A multi-agent system is an AI architecture in which several language-model agents, each with its own prompt, tools, and context, work on parts of a task and pass results to one another.

A multi-agent system is an AI architecture in which several AI agents work on parts of one task and exchange results. Each agent has its own instructions, tools, and context window. The agents coordinate through an orchestrator agent, a shared state store, or a protocol such as A2A. The term is older than language models and also names a field of distributed AI research; in LLM engineering it usually means a group of LLM-driven agents.

How a multi-agent system works#

A multi-agent system usually works through an orchestrator-worker pattern: a lead agent splits the task, starts subagents with narrower instructions, and merges what they return. Each subagent runs its own loop in a separate context, so it can explore one branch of the problem without filling the lead agent's context with intermediate results.

A lead agent splits a task across three subagents, each working in its own context, and their condensed findings return to the lead to be merged. Lead agentPlans and splits the taskSubagent AOwn contextSubagent BOwn contextSubagent COwn contextLead merges condensed findings
Figure 1: The orchestrator-worker pattern in a multi-agent system.

Anthropic described this design for its multi-agent research system: a lead agent plans the research and spawns subagents that search in parallel, then a separate agent attributes claims to sources. Other topologies include fixed pipelines, where each agent hands off to the next, and peer networks, where agents discover each other through a protocol.

Multi-agent system vs single agent#

A multi-agent system outperforms a single agent on tasks that split into independent parallel parts, and it underperforms on tasks with tight dependencies between steps. On Anthropic's internal research evaluation, a multi-agent setup with Claude Opus 4 as lead and Claude Sonnet 4 subagents outperformed a single-agent Claude Opus 4 by 90.2%.

The same report lists the costs. Multi-agent systems used about 15 times the tokens of a chat interaction, and Anthropic called most coding tasks a poor fit because they offer few parallelizable subtasks. A practical rule: start with one agent, and add agents only when a measured bottleneck, such as context size, parallel search, or separation of permissions, justifies the extra cost and debugging surface.

See what AI agents are, the AI agent frameworks that support multi-agent orchestration, and A2A vs MCP for the protocols agents use to communicate.

Cite this definition

The AI Engineering Editors. "What is a multi-agent system in AI?" The AI Engineering, updated Nov 06, 2026. https://theaiengineering.com/glossary/multi-agent-system

Title
What is a multi-agent system in AI?
Author
The AI Engineering Editors
Publisher
The AI Engineering
Updated
View as Markdown

Part of the AI agents hub · Updated