A multi-agent system is a setup where several AI agents work on one job together, each taking a part, and either pass work to each other or report to a lead agent that puts the results together. It can do better than one agent on wide jobs that split into parts, but it uses far more computing and has more ways to go wrong.
Think of a newsroom on election night. The editor hands each reporter one region, the reporters work at the same time, and the editor writes the front page from what they file. One reporter could cover it all, just much more slowly. A multi-agent system is that newsroom, with AI agents as the editor and the reporters.
In a sentence: “The research tool is a multi-agent system: a lead agent splits the question, and helper agents search different angles at once.”
What counts as a multi-agent system?
The plain point: more than one AI agent, sharing one goal.
An AI agent is an AI that plans and takes steps on its own, often using tools. IBM’s definition of the group version: “A multi-agent system (MAS) consists of multiple artificial intelligence (AI) agents working collectively to perform tasks”, on behalf of a person or another system.1 Anthropic, which makes Claude, puts it in one line: “A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together.”2
How is it different from similar terms?
The plain point: the words overlap; the number of agents is what sets this one apart.
| Term | How it differs |
|---|---|
| Single agent | Differs: one AI does every step itself. Simpler to follow, slower on wide jobs. |
| Subagent | Differs: one helper inside a multi-agent setup, sent off by a lead agent and reporting back. |
| Agentic AI | Differs: AI that works towards a goal with little supervision. It can be one agent or many. |
| Prompt chaining | Differs: a fixed series of prompts, each feeding the next. The steps are set in advance, not chosen by agents. |
How are multi-agent systems organised?
The plain point: either a boss and workers, or peers that talk to each other.
The common AI pattern today is a lead agent with helpers. Anthropic’s research feature uses “an orchestrator-worker pattern”, in which “a lead agent coordinates the process while delegating to specialized subagents that operate in parallel.”2 IBM describes the two broad shapes: “In centralized networks, a central unit contains the global knowledge base”, while “Agents in decentralized networks share information with their neighboring agents instead of a global knowledge base.” It notes the central unit is a single point the whole system depends on.3
When does a multi-agent system help, and when not?
The plain point: it helps when the work splits into parts that can run at once; it costs a lot either way.
Anthropic reported that its system, with one lead agent and helper agents, “outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval.” That is the company’s own test of its own product. The same post says multi-agent systems “excel at valuable tasks that involve heavy parallelization” and “use about 15× more tokens than chats”, and that “most coding tasks involve fewer truly parallelizable tasks than research”.2 Early versions went wrong in ways a single agent would not, such as “spawning 50 subagents for simple queries”.4
Outside researchers are more cautious. A study of popular multi-agent frameworks found “14 unique modes, clustered into 3 categories” of failure, “(i) system design issues, (ii) inter-agent misalignment, and (iii) task verification”, and said their “performance gains on popular benchmarks are often minimal.”5 Anthropic’s general advice on agents points the same way: “we recommend finding the simplest solution possible, and only increasing complexity when needed.”6
Try one agent first. Reach for several when the job splits into parts that can run at the same time, need different tools, or need a separate checker, and expect it to use much more computing. This site’s multi-agent guide has a test to run before you build one.
Questions people ask
Is a multi-agent system better than one agent?
Sometimes. Anthropic measured a large gain on wide research questions, while an outside study of multi-agent frameworks found gains on common benchmarks “often minimal”.5 It depends on whether the work truly splits.
Does giving each agent a job title help?
Not by itself. A study of personas in system prompts found that adding them “does not improve model performance across a range of questions”; splitting agents by tools or permissions means more than splitting by titles. See role prompting.7
What is the orchestrator in a multi-agent system?
The part that splits the goal into tasks, hands them out and combines the results; in Anthropic’s system it is the lead agent that coordinates while “delegating to specialized subagents”.2 What is an orchestrator? covers the word on its own.
Where to go next
- Before you build one: multi-agent: when it helps and when it does not.
- The helpers inside it: what is a subagent?
- One agent first: what is an AI agent? and agentic AI.
- Finding where a run went wrong: diagnosing a run.
- The words: the AI glossary.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is a multi-agent system?
- One AI model trained on many languages
- Several AI agents working on one job together
- A chatbot that remembers past conversations
Show the answer
Answer: Several AI agents working on one job together. The page quotes IBM and Anthropic: multiple agents working together or collectively on tasks.
2. According to Anthropic, what kind of task suits a multi-agent system?
- Tasks that split into many parts that can run at the same time
- Short questions with one right answer
- Tasks where every step depends on the one before
Show the answer
Answer: Tasks that split into many parts that can run at the same time. Anthropic says multi-agent systems excel at tasks that involve heavy parallelization, and that most coding tasks have fewer parallel parts than research.
3. What does the page suggest trying before building a multi-agent system?
- Giving each agent a more impressive job title
- Adding as many agents as possible
- One agent first, adding more only when the work truly splits
Show the answer
Answer: One agent first, adding more only when the work truly splits. The page cites Anthropic’s advice to find the simplest solution and a study finding gains on benchmarks often minimal.
Related terms: Function calling · Human-in-the-loop · MCP (Model Context Protocol) · Orchestrator · Plugin · RAG
- 1IBM, What is a Multi-Agent System?, IBM Think, undated, read at source 19 Oct 2026“A multi-agent system (MAS) consists of multiple artificial intelligence (AI) agents working collectively to perform tasks”. ↑
- 2Anthropic, How we built our multi-agent research system, 13 Jun 2025, read at source 19 Oct 2026“A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together.” ↑a“Our Research system uses a multi-agent architecture with an orchestrator-worker pattern, where a lead agent coordinates the process while delegating to specialized subagents that operate in parallel.” ↑b“outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval”; “multi-agent systems excel at valuable tasks that involve heavy parallelization”; “multi-agent systems use about 15× more tokens than chats”; “most coding tasks involve fewer truly parallelizable tasks than research”. ↑c“a lead agent coordinates the process while delegating to specialized subagents that operate in parallel.” ↑d
- 3IBM, What is a Multi-Agent System?, read at source 19 Oct 2026“In centralized networks, a central unit contains the global knowledge base” and “Agents in decentralized networks share information with their neighboring agents instead of a global knowledge base.” ↑
- 4Anthropic, How we built our multi-agent research system, read at source 19 Oct 2026early agents made errors like “spawning 50 subagents for simple queries”. ↑
- 5Cemri, Pan, Yang et al., Why Do Multi-Agent LLM Systems Fail?, arXiv v3, 26 Oct 2025, read at source 19 Oct 2026“14 unique modes, clustered into 3 categories”; “(i) system design issues, (ii) inter-agent misalignment, and (iii) task verification”; “their performance gains on popular benchmarks are often minimal.” ↑a“their performance gains on popular benchmarks are often minimal.” ↑b
- 6Anthropic, Building effective agents, 19 Dec 2024, read at source 12 Oct 2026 (as cited on this site’s AI agent page): “we recommend finding the simplest solution possible, and only increasing complexity when needed.”↑
- 7Zheng, Pei, Logeswaran, Lee & Jurgens, When “A Helpful Assistant” Is Not Really Helpful, Findings of EMNLP 2024, read at source 19 Oct 2026“adding personas in system prompts does not improve model performance across a range of questions”. ↑
Cemri, Pan, Yang et al., Why Do Multi-Agent LLM Systems Fail?, arXiv v3, 26 Oct 2025, read at source 19 Oct 2026.
Zheng et al., When “A Helpful Assistant” Is Not Really Helpful, Findings of EMNLP 2024, read at source 19 Oct 2026.
The comparisons and examples on this page (the election-night newsroom, the research tool sentence) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK