Multi-agent systems (MAS)

A multi-agent system is a setup in which several AI agents work on one task together, each doing its own part. Ask it to compare twenty suppliers, and a lead agent can give each helper agent a few suppliers to research at the same time, then combine their findings into one overview. That helps with large tasks that split into independent pieces, and every extra agent adds to the cost.

Where does the term come from?

The idea is older than language models. In computer science a multi-agent system is any system of several interacting agents, used for robots, traffic control and logistics, among other things. With large language models it became a research area of its own, in which each agent is an AI agent built on a large language model: in Anthropic's words, an LLM "autonomously using tools in a loop".

How does it work?

A common structure is orchestrator-worker. In Anthropic's Research feature (How we built our multi-agent research system, June 2025), a lead agent plans the research and starts 3 to 5 subagents that each search one part of the question in parallel, each in its own context window. The lead agent combines their condensed results. Because each subagent has its own tools, prompt and search path, the search depends less on one early choice. Other structures, such as a handoff to a specialist agent or one agent reviewing another's output, are covered under AI orchestration.

Anthropic's advisor tool (beta, April 2026) turns the usual order around. A cheaper model such as Haiku does the work and consults Opus only on decisions it cannot handle; Opus returns a plan, a correction or a stop signal and never calls tools itself. On BrowseComp, a benchmark for finding hard-to-find information on the web, Haiku with an Opus advisor scored 41.2%, against 19.7% alone. Routing can also be learned: AMRO-S (March 2026) assigns requests to agents with ant colony optimisation and reports a better balance of quality and cost than other routing methods on five public benchmarks.

When is it worth it?

When a task splits into parts that can be done at the same time. In Anthropic's internal research evaluation, Claude Opus 4 as lead agent with Claude Sonnet 4 subagents performed 90.2% better than Claude Opus 4 on its own. Asked for the board members of all IT companies in the S&P 500, the multi-agent system split the question into tasks for subagents, while the single agent searched step by step and did not find the answer. The price is tokens: in Anthropic's data an agent uses about 4 times as many as a chat, and a multi-agent system about 15 times. Tasks in which all agents need the same context, or depend on each other's steps, suit it less; Anthropic names most coding tasks.

Where does it go wrong?

Cemri et al. (Why Do Multi-Agent LLM Systems Fail?, 2025) found that multi-agent systems often gain little on popular benchmarks. In more than 1,600 annotated runs of seven frameworks they found 14 failure modes in three groups: flaws in the system design, agents that ignore or misread each other, and missing or weak checks of the result. Anthropic's early research agents started 50 subagents for simple questions, and small changes to the lead agent changed the subagents' behaviour in ways nobody predicted. Oversight gets harder too: work spread over parallel subagents no longer reads as one thread of actions that a user can follow and steer (Trustworthy agents in practice, April 2026).

Example

In a March 2026 study by Ahtisham, Vanacore and Kizilcec, models from OpenAI, Anthropic and Google labelled 800 teacher utterances from 63 transcripts of maths lessons. Each model then checked its own label against the rubric, and an independent adjudicator model settled disagreements. These two steps improved the match with the experts' labels in all seven categories. With the adjudicator, models without extended reasoning came close to reasoning models (an F1 score of about 0.60 against 0.61) with 20 to 25% fewer tokens.