Multi-agent systems with Claude: when to use one, and how the coordinator pattern works
A multi-agent system splits one job across several AI agents: typically a coordinator (lead agent) that plans and delegates, and subagents that each work on one part in their own context, often in parallel. It shines on broad, parallel work such as research, and wastes tokens on tightly linked work such as most coding. The coordinator's delegation instructions decide whether it works.
What is a multi-agent system?
A multi-agent system is several AI agents working on one task. The common shape with Claude is orchestrator and workers: Anthropic's own research system uses "a lead agent [that] coordinates the process while delegating to specialized subagents", and the subagents run in parallel.
Each subagent is a full agent with its own context window, its own instructions and its own tools. It explores its part, then returns a condensed result to the lead, which combines the pieces. With Claude you build this with subagents, in Claude Code or the Claude Agent SDK.
How the coordinator pattern works
- Plan. The coordinator reads the task and decides how to split it.
- Delegate. It starts subagents, each with one clearly bounded part, often several at once.
- Work in isolation. Each subagent searches, reads and reasons in its own context; its intermediate steps never reach the coordinator.
- Report. Each returns a short, structured result.
- Combine and check. The coordinator merges results, spots gaps and may send another round.
For a bike-rental support team, a coordinator could hand a complex complaint to three subagents: one checks the booking history, one reviews the damage report, one drafts the refund summary, and the coordinator writes the reply.
When a multi-agent system beats a single agent
Anthropic reports that its multi-agent research system, with a lead agent and parallel subagents, "outperformed single-agent Claude Opus 4 by 90.2%" on its internal research evaluation. The reason is mostly capacity: on their browsing benchmark, "token usage by itself explains 80% of the variance" in performance, and several agents can spend far more tokens on a problem than one context window allows.
Multi-agent fits work that is:
- broad: many sources or directions to explore;
- parallel: parts that don't depend on each other;
- too big for one context window.
When it is the wrong choice
The same write-up is clear about the limits. Domains that "require all agents to share the same context" fit poorly, and "most coding tasks involve fewer truly parallelizable tasks than research". When every step depends on the last, splitting the work adds coordination cost and lost context without adding capacity.
And it is expensive: "agents typically use about 4× more tokens than chat interactions," and "multi-agent systems use about 15× more tokens than chats." Use one where the value of the answer pays for that.
How to delegate well: the coordinator's instructions
Most multi-agent failures are delegation failures. Anthropic's rule: "Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries."
| Give each subagent | Why |
|---|---|
| A clear objective | Vague tasks get duplicated or misread |
| An output format | So results can be merged |
| The tools and sources to use | So it doesn't waste turns exploring |
| Boundaries | So two subagents don't do the same work |
| The context it needs | A subagent starts fresh: it only knows what the prompt tells it |
Two more lessons from the same system: scale effort to the query (a simple fact-check needs one agent and a few calls, not ten subagents), and start wide, then narrow (short, broad searches first, then focused ones). The lead agent also spins up 3 to 5 subagents in parallel rather than one after another.
Errors, testing and production
Multi-agent systems are harder to run. "Agents are stateful and errors compound": a subagent that silently swallows a failure poisons the coordinator's answer, so subagents should report what they tried, what failed and what they found, rather than an empty "no results". Runs are non-deterministic, so debugging needs good tracing. For evaluation, Anthropic starts with a few examples right away and uses an LLM judge with a clear rubric once the system scales.
Multi-agent systems on the Claude Certified Architect exam
Domain 1 (27%) of the CCAR-F exam covers coordinator and subagent orchestration (task 1.2), passing context to subagents (1.3) and task decomposition (1.6), and Domain 5 adds error propagation across agents. Scenario 3, the multi-agent research system, is built entirely on this pattern. The traps to expect: assuming subagents share the coordinator's context, splitting work that is not parallel, and letting a subagent hide a failure.
More: Claude subagents, Claude Agent SDK and all articles.
Sources: How we built our multi-agent research system (Anthropic Engineering); Building effective agents; Subagents in the Agent SDK; Claude Certified Architect Foundations exam guide v1.0, task statements 1.2, 1.3, 1.6 and 5.3.
FAQ
What is a multi-agent system in AI?
A system where several AI agents work on one task: usually a lead agent that plans and delegates, and specialized subagents that each handle a part, in their own context, and report back.
When is a multi-agent system better than a single agent?
When the work splits into independent parts that can run in parallel and would overflow one context window, such as broad research. Anthropic reports its multi-agent research system beat a single agent by 90.2% on an internal research eval.
Do multi-agent systems cost more?
Yes. Anthropic measured agents using about 4 times more tokens than chat, and multi-agent systems about 15 times more. Use them where the task is valuable enough to pay for that.
How do I build a multi-agent system with Claude?
With the Claude Agent SDK or Claude Code: define subagents with a description, a prompt and their tools, and let the main agent delegate through the Agent tool, passing each one the context it needs.