Creativity · Agent Protocol
Claude Subagents in Production
Anthropic's own multi-agent research system — the engine behind Claude's web-research capability — uses the same subagent primitive developers get in Claude Code: a lead Claude plans, dispatches focused sub-Claudes in parallel, and integrates their findings. Anthropic has published detailed engineering notes showing that well-tuned multi-agent Claude significantly outperforms a single Claude on complex research tasks, despite higher token costs.
Protocol facts
- Sponsor
- Anthropic
- Status
- stable
- Spec
- https://www.anthropic.com/engineering/multi-agent-research-system
- Interop with
- Claude Code, MCP, A2A, Claude Agent SDK
Frequently asked questions
Why does multi-agent Claude outperform single Claude for research?
Anthropic's analysis attributes the lift primarily to parallel exploration — multiple sub-Claudes can cover more of the search space concurrently than one can in sequence — plus narrower per-agent context keeping each sub-Claude focused.
What's the cost trade-off?
Multi-agent systems use noticeably more tokens — often several times more than a single agent. Anthropic's guidance is to reserve the pattern for high-value tasks where the quality gain is worth the spend.
Does this use A2A over the wire?
Today these are in-process sub-Claudes. But the pattern maps cleanly onto A2A — the lead could dispatch to remote agents just as easily, which is where the broader ecosystem is heading.
Sources
- Anthropic — How we built our multi-agent research system — accessed 2026-04-20