Key takeaways

  • A multi-agent system has several LLM agents work on one task, usually with a lead agent that splits the work and merges the results.
  • Anthropic says its multi-agent research system beat a single Claude Opus 4 agent by 90.2% on its internal eval, while using about 15 times the tokens of a chat.
  • A 2025 study of more than 1,600 traces found 14 ways multi-agent systems fail, in system design, misalignment between agents and task verification.

A multi-agent system is an AI setup where several language-model agents work on one task together, usually with a lead agent that breaks the job into parts, hands each part to another agent and combines what comes back. Each agent has its own instructions, tools and context, so the work can run in parallel and no single agent has to hold everything at once. It costs more and fails in new ways, which is why most tasks still run on one agent.

The term is in the news because Perplexity launched Multiagent on November 17, 2026, a new effort level in Perplexity Computer. A lead agent plans the task and can split it across up to 16 agents running different frontier models.

What a multi-agent system is

Start with one agent. An AI agent is a language model running in a loop: it reads the task, calls a tool such as web search or a code runner, reads the result and decides what to do next, until it is done. The software that runs that loop, holds the tools and manages memory is the agent harness.

A multi-agent system puts several of those loops together. The agents pass work along, talk to each other or report to a coordinator, and they can run the same model or different ones.

How multi-agent systems work: the common patterns

Pattern How it works Example
Orchestrator and workers (also "supervisor" or "subagents") A lead agent plans, sends subtasks to worker agents and merges their results. Workers report only to the lead. Anthropic's Research feature
Handoffs One agent passes the conversation to a specialist, which takes over. A triage agent routing a customer to a billing agent
Router A first step classifies the request, sends it to one or more specialists and combines the answers. A help desk that routes by topic
Team with shared discussion Agents message each other, check each other's work and flag disagreements. Perplexity's Multiagent
Role-based assembly line Agents play job roles in a fixed process, each passing its output to the next. MetaGPT and ChatDev for software

The OpenAI Agents SDK offers two building blocks. With "agents as tools", a manager agent keeps the conversation and calls specialists for bounded subtasks. With a handoff, the specialist takes over for the rest of the turn. LangChain's multi-agent docs, which cover LangGraph, call the core design problem context engineering: choosing what each agent gets to see.

Three 2023 research frameworks set many early patterns: AutoGen, built on agents that converse with each other, and MetaGPT and ChatDev, which give agents software-team roles.

A real example: Anthropic's research system

Anthropic's engineering post on its multi-agent research system, published in June 2025, describes the design behind the Research feature in Claude.

A lead agent, Claude Opus 4 at the time, reads the question, makes a plan and saves it to memory. It starts subagents running Claude Sonnet 4, usually three to five at once, each with its own context window and search tools. Each one explores an angle and sends back a condensed result.

Anthropic reports these results from its own tests:

  • The multi-agent setup "outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval."
  • Running subagents and their tool calls in parallel cut research time by up to 90% for complex queries.
  • On the BrowseComp browsing benchmark, token usage alone explained 80% of the variance in performance. Tool calls and model choice explained most of the rest.

That last point explains the approach: splitting a job lets the system spend more tokens than one context window holds, and for research, more tokens buy better answers.

Perplexity's Multiagent, launched November 17

Perplexity's version adds cross-checking between models. A lead agent running Claude Fable 5.2 divides the task among agents that run, by default, Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. GPT 6 Astra and Fable 5.2 join when the user asks or the task calls for them. The lead can assign up to 16 agents and does part of the work itself.

The agents share findings directly and flag discrepancies. In Perplexity's Ultra level, subagents report only to the lead. Multiagent is available to Perplexity Max subscribers on the web. The post gives no benchmark results and no credit cost.

When a multi-agent system beats a single agent

It tends to pay off when:

  • The work splits into independent parts. Research is the clearest case: several subagents can read different sources at once.
  • There is more material than one context window holds. Each agent works through its share in a fresh context and returns a summary.
  • One agent has too many tools. LangChain's docs say splitting helps when a single agent picks the wrong tools from a long list.
  • The work needs checking. Separate agents, or different models as in Perplexity's design, can review each other.

It tends not to when:

  • Every agent needs the same context, or the steps depend heavily on each other. Anthropic names both as poor fits.
  • The task is coding. Anthropic says "most coding tasks involve fewer truly parallelizable tasks than research," and that agents still struggle to coordinate with each other in real time.
  • The question is simple. LangChain's docs advise against multi-agent by default, since one agent with the right tools and prompt can often get similar results.

What multi-agent systems cost

Tokens are the main cost. Anthropic says agents typically use about 4 times as many tokens as a chat, and multi-agent systems about 15 times as many. On a pay-per-token API, the bill grows with that. Anthropic's conclusion is that the task has to be valuable enough to pay for the extra work.

Why multi-agent systems fail

The paper "Why Do Multi-Agent LLM Systems Fail?", by Mert Cemri, Melissa Z. Pan, Shuyi Yang and colleagues, starts from an awkward fact: on popular benchmarks, multi-agent systems often gain little over single agents. The authors annotated more than 1,600 traces from seven frameworks, including MetaGPT, ChatDev and Magentic-One, with human annotators agreeing at a kappa of 0.88. Their taxonomy, MAST, lists 14 failure modes in three groups:

  • System design issues: agents that disobey the task or their role, repeat steps, lose the conversation history or don't know when to stop.
  • Misalignment between agents: agents that fail to ask for clarification, drift off the task, withhold information, ignore another agent's input, or say one thing and do another.
  • Task verification: agents that stop too early, or checks that are missing, incomplete or wrong.

Anthropic saw the same kinds of problems. Its agents spawned 50 subagents for simple queries and searched endlessly for sources that didn't exist. Vague instructions led to duplicated work. The fixes were plain: give each subagent a clear objective, output format, tools and limits, and scale the team to the question, from one agent with 3 to 10 tool calls for a simple fact to more than 10 subagents for complex research.

Where this leaves multi-agent systems

A multi-agent system trades money for reach. On wide research tasks the published numbers support it. On tightly linked work like most coding, one well-equipped agent is usually cheaper. Anthropic's headline results come from its own tests, and Perplexity has published none.

Frequently asked questions

What is the difference between a single agent and a multi-agent system?

A single agent is one model running in a loop with tools until the task is done. A multi-agent system runs several such agents, each with its own instructions and context, and coordinates them, usually through a lead agent.

What are examples of multi-agent systems?

Anthropic's Research feature in Claude uses a lead agent with parallel search subagents, and Perplexity's Multiagent mode, launched on November 17, 2026, splits a task across up to 16 agents running different models. Research frameworks such as AutoGen, MetaGPT and ChatDev build agent teams for coding and other tasks.

Are multi-agent systems better than a single agent?

On wide tasks that split into independent parts, such as research, they can be: Anthropic reports a 90.2% gain on its own eval. On tightly linked work such as most coding, the gains are often small, and they cost far more tokens.

What frameworks are used to build multi-agent systems?

Common choices include the OpenAI Agents SDK, which supports handoffs and agents as tools, and LangGraph from LangChain, which supports subagents, handoffs and routers. Microsoft's original AutoGen repository is now archived.

Sources

  1. How we built our multi-agent research systemAnthropic
  2. Introducing Multiagent for the most complex workPerplexity
  3. Why Do Multi-Agent LLM Systems Fail?arXiv
  4. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent ConversationarXiv
  5. MetaGPT: Meta Programming for A Multi-Agent Collaborative FrameworkarXiv
  6. ChatDev: Communicative Agents for Software DevelopmentarXiv
  7. Orchestrating multiple agentsOpenAI
  8. Multi-agentLangChain

Add ClaudeGPT Si as a preferred source on Google