Multi Agent Systems: What Anthropic and Cognition Found

Multi agent systems are groups of AI agents that split a job between them, each with its own instructions, tools and working memory. IBM describes them as agents that collaborate to reach shared, system-wide goals. The idea is real and in production, but the evidence for it is narrower than the marketing around "autonomous AI agents replacing software" suggests.

The best public numbers come from two companies that disagree with each other. Anthropic reports a large gain from a multi-agent research feature. Cognition, the maker of the coding agent Devin, argued in 2025 that most multi-agent designs are fragile, then wrote in 2026 about the setups that do work.

What are multi agent systems?

An AI agent is a language model that runs in a loop: it reads a goal, picks a tool (a web search, a code editor, a database query), looks at the result and decides what to do next. A multi agent system puts several of those loops side by side and has them hand work to one another.

IBM lists three common layouts. In a centralized layout, one unit holds the shared knowledge and connects the agents, which makes communication easy but means a failure there can take everything down. In a decentralized layout, agents trade information with their neighbours, which is more robust but harder to coordinate. In a hierarchical layout, agents sit in a tree with different levels of authority.

The practical appeal is specialization and scale. One agent can search the web while another reads documents and a third checks the answer, each with a fresh, uncluttered context (the amount of text a model can hold in mind at once).

How does a multi-agent setup work in practice?

Anthropic's engineering team published the most detailed public description, for the Research feature in its Claude apps. A lead agent reads the user's question, writes a plan, and saves that plan to memory so it survives if the conversation outgrows the model's context limit. It then starts several subagents, each with its own context window, tools and sub-task. Anthropic says the lead typically spins up three to five of them in parallel.

The subagents search, then return what they found. The lead agent combines the findings and decides whether more research is needed. A separate citation agent finally attaches sources to the claims in the answer.

This pattern is usually called orchestrator-worker. The orchestrator plans and merges; the workers do the reading and searching, and each returns only its findings to the lead agent. Anthropic says parallel work cut research time by up to 90% for complex queries.

Is a multi-agent setup better than a single agent?

On the right kind of task, according to Anthropic. It reports that a lead agent running Claude Opus 4 with Claude Sonnet 4 subagents beat a single Claude Opus 4 agent by 90.2% on its internal research evaluation. That is Anthropic's own test on its own product, so it shows the approach can work and says little about other tasks.

The same post attributes much of the gain to token spending. On a web-browsing benchmark called BrowseComp, Anthropic found that token usage alone explained about 80% of the differences in performance. A token is a small chunk of text, and the more tokens an agent reads and writes, the more it can look at. Multiple agents are a way to spend more tokens on a problem than one agent's context window allows.

That spending is large. Anthropic reports that agents use roughly four times the tokens of an ordinary chat, and multi-agent systems roughly fifteen times. Anthropic says the task has to be valuable enough to justify that cost, which rules out low-value jobs.

Why did Cognition tell people not to build multi-agent setups?

In June 2025, Cognition engineer Walden Yan published "Don't Build Multi-Agents." It gave two principles: share full context between agents (their complete traces), and remember that every action carries implicit decisions that can conflict with another agent's.

His example was a request to build a Flappy Bird clone. The work was split into two sub-tasks: a game background, and a bird. One subagent produced a Super Mario style background, the other a bird that looked and moved nothing like Flappy Bird's. A final agent then had to reconcile the mismatch. Even with more shared context, the post says, the two kept making conflicting assumptions about visual style that nobody had specified.

Yan's recommendation was a single-threaded agent, optionally with a model that compresses the history for long jobs. Anthropic's own list of weak spots overlaps with his: tasks where agents need the same context or depend heavily on each other, and most coding tasks.

Which multi agent setups work today?

Cognition revisited the question in a post dated April 22, 2026, "Multi-Agents: What's Actually Working." Its main finding is that multi-agent systems work when only one agent writes at a time and the extra agents contribute judgement instead of actions.

It names four patterns that work. The first is a code-review loop, where one agent writes code and a second reviews it with a clean context. Cognition says Devin Review catches an average of two bugs per pull request (a proposed code change), about 58% of them severe.

The second is a "smart friend" setup, where a smaller model asks a larger one for help on hard problems. The third is a manager that splits a job among child agents and merges the results. The fourth is read-only subagents that search the web or a codebase and change nothing.

What did not work: Cognition calls unstructured swarms, where arbitrary agents negotiate with each other, "mostly a distraction." It also reports that a smart-friend setup failed when the primary model was too far behind the helper to know when to ask. The open problems it lists are about communication: when a weaker model should escalate, how child agents share discoveries with siblings, and how to pass context without overloading the receiver.

Both companies find parallel agents that read and report useful, while several agents all editing the same thing at once remain risky, because each makes its own assumptions about the work.

Are AI agents replacing SaaS software?

Nothing in the Anthropic or Cognition posts shows that. They describe agents that call existing tools, such as web search, code search and code review.

Tool access is the layer that makes that possible. Standards such as the Model Context Protocol, covered in our explainer on MCP in production, give agents a common way to reach those tools. Whether that adds up to agents replacing dashboards is a prediction, and the public numbers do not settle it.

Gartner's forecast, published in June 2025, leans the other way. It predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls. That is an analyst's projection, and it matches the cost and reliability themes in the engineering posts.

What to watch next

Anthropic lists problems that remain open. Errors compound in long-running agents, so it uses checkpoints and retries. Runs are non-deterministic, which makes failures hard to diagnose. Lead agents currently wait for each batch of subagents to finish, and running them asynchronously would help speed but add complexity. Small changes to the lead agent can also change subagent behaviour in ways nobody predicted.

For a product that claims to use many agents, ask whether the task splits into independent pieces or the pieces depend on each other, whether a correct answer is worth roughly fifteen times the token bill, and whether a single agent makes the final edits with a reviewer using clean context.

Anthropic says early versions of its system searched endlessly and spawned too many subagents for simple queries. Our piece on stopping infinite loops in AI agent workflows covers one way teams constrain that.

Cognition says enterprise use of Devin grew about eightfold over the prior six months, while swarms of agents that all write changes still see little meaningful adoption. If that pattern continues, the designs that spread will be the plain ones: a clear planner, a few read-only helpers and a reviewer at the end.