Multi-agent system
Several specialized AI agents working together to solve a complex task.
Explanation
A multi-agent system consists of several AI agents with their own roles, tools and instructions that work on a task together. Often an orchestrator agent coordinates specialized worker agents.
How it works
An orchestrator breaks the task into subtasks and assigns them to specialized agents, for example for research, analysis or writing. The workers operate, often in parallel, with their own context and return results. The orchestrator reviews, combines and decides on next steps.
Example
For a market report, several worker agents research individual competitors in parallel, an analysis agent condenses the findings, a review agent checks sources and figures, and the orchestrator compiles the final report.
Why it matters
Single agents hit limits in context, speed and quality on large, multi-layered tasks. Splitting the work among specialized agents makes such tasks manageable and can lower costs when workers use cheaper models.
Orchestrator-worker and other architectures
The most common pattern is orchestrator-worker: a central agent plans and delegates, specialized agents execute. There are also sequential pipelines, in which each agent handles a fixed step and passes the result on, and reviewer patterns, in which one agent critically evaluates another agent's output and sends it back until it meets the quality criteria.
The economic benefit comes from not giving every agent the most expensive model. A strong model plans and reviews, while cheaper models handle the subtasks. How much this can save is shown in our article on the Advisor and Orchestrator Pattern.
When is a multi-agent system worth it?
Multi-agent systems shine when a task can be split into independent parts that can be worked on in parallel, or when the context becomes too large for a single agent. Examples are broad research, creating extensive documents or processing many similar cases.
The price is complexity. More agents mean more model calls, more potential failure points at the handoffs and more demanding tests. Anthropic reports that its own multi-agent research system uses about 15 times as many tokens as a regular chat. For many tasks, a single, well-equipped agent is therefore the better choice.
Frequently asked questions
What is a multi-agent system?
A multi-agent system is an AI system in which several agents with different roles work together, usually coordinated by an orchestrator, to solve a task that would be too large for a single agent.
What is the difference between an AI agent and a multi-agent system?
A single agent plans and acts on its own in a loop. A multi-agent system distributes the work across several agents with their own context, which exchange results and build on each other.
Which frameworks are used to build multi-agent systems?
Common choices include LangGraph and CrewAI as well as the agent SDKs from major providers such as OpenAI, Anthropic and Microsoft. Testability, monitoring and integration usually matter more for success than the framework itself.