Multi-Agent System
Some tasks get handled by one agent from start to finish. Others get split across several agents that each work on a piece, then combine what they find. A multi-agent system is the second kind.
You've already seen one way to build one: an agent delegating to a subagent. A multi-agent system is the broader idea. It covers a single delegation, several agents working at once, or a structured pattern called orchestrator-workers, where one lead agent breaks a task into pieces, hands each piece to a separate worker agent, and combines what comes back.
Why split a task across agents
The main reason is speed through parallel work. If a task breaks into independent pieces, several agents can investigate at the same time instead of one agent working through everything in sequence. A second reason is keeping each piece's search process separate. In a single agent, every search and every result piles into the same context — the pool of text the model can see when deciding what to do next. Splitting the work across agents means each one explores its own piece in its own separate context, instead of cluttering one shared pool with details that don't matter to the others.
Anthropic's research system
Anthropic's own research system ran into this directly. Asked to find every board member of every company in the Information Technology sector of the S&P 500, a single agent worked through the list one company at a time. Its searches were slow and sequential, and it failed to complete the task. A multi-agent version split the work into one search per company, ran those searches in parallel across several subagents, and found the correct answer. On Anthropic's internal research evaluation, a lead agent coordinating parallel subagents outperformed a single agent by over 90%. That number is specific to their system and that kind of task — many independent lookups explored at once — not a universal guarantee.
The real cost
This doesn't come free. By Anthropic's own accounting, a single agent typically uses around four times the tokens of a normal chat reply. A multi-agent system uses around fifteen times the tokens of that same normal chat reply — not fifteen times the single-agent figure.
When this is worth it
When a task genuinely decomposes into independent pieces, and finishing faster or covering more ground is worth paying for. Research that fans out across many independent sources is the clearest case — each source can be checked in parallel with no piece depending on another's result.
When it isn't
When a single agent could do the job directly, or the pieces aren't actually independent. If each part needs the result of the last one, running them in parallel doesn't help — it just adds coordination cost for nothing. It's also not worth it when the task's value doesn't clear that roughly fifteen-times cost bar. A quick task with a low-stakes answer rarely justifies a multi-agent setup, no matter how parallelizable it looks on paper.
In this guide
FAQ
What happens if two agents in the system produce conflicting results?
Something has to reconcile them — usually a lead agent that combines the pieces, or a separate synthesis step that resolves conflicts before producing a final answer. That reconciliation step is itself extra work, and it's part of why multi-agent systems are harder to get right than they look on paper.