2.9 KiB
How we built our multi-agent research system (Anthropic Engineering)
Metadata
| Field | Value |
|---|---|
| Institute | Molodetz Institute — Research & Retrieval Division |
| Subject | cap-protocol |
| Document type | Source Dossier |
| Tier | 1 |
| Source title | How we built our multi-agent research system (Anthropic Engineering) |
| Source URL | https://www.anthropic.com/engineering/multi-agent-research-system |
| Source publisher / author | Anthropic |
| Source type | blog post |
| Source date (published) | Unknown — not stated |
| Date researched | 2026-10-25 |
| Researched by | tai (agent), model: Claude Sonnet 5, on behalf of retoor |
| Research method | direct page fetch |
| Confidence | High |
| Status | Active |
Summary
Anthropic's account of an orchestrator-worker research system: a lead agent plans and spawns parallel subagents with their own context windows. Includes numbers on performance and token cost, prompting principles for delegation, and when the pattern does not fit.
Key findings
- Lead Claude Opus 4 with Sonnet 4 subagents beat single-agent Opus 4 by 90.2% on the internal research eval.
- On BrowseComp three factors explained 95% of performance variance; token usage alone explained 80%, with tool calls and model choice the others.
- Agents use about 4x the tokens of chat; multi-agent systems about 15x. Parallelization cut research time by up to 90% on complex queries.
- Delegation must be taught: each subagent needs an objective, an output format, tool and source guidance, and clear task boundaries; vague instructions caused duplicated work.
- Scale effort to complexity: 1 agent with 3-10 tool calls for simple facts, 2-4 subagents with 10-15 calls for comparisons, 10+ subagents for complex research.
- Does not fit: tasks where all agents share the same context or have many dependencies, most coding tasks (fewer parallelizable parts), weak real-time coordination, and low-value tasks that cannot pay the token cost.
- Production: agents are stateful so errors compound (resume from checkpoints, tell agents when a tool fails); leads wait synchronously for each subagent batch; start evals with about 20 real queries; judge end state for stateful agents.
- Upgrading to Sonnet 4 gave a larger gain than doubling the token budget on Sonnet 3.7.
Pros / strengths
Primary source with concrete numbers, and an honest list of where delegation fails, including coding.
Cons / limitations / caveats
A research workload, not an interactive coding agent; numbers come from Anthropic's internal eval and BrowseComp, not independently reproduced. 15x is a ballpark for their system.
Relevance to tai
Direct evidence for two rules: delegation to another model costs a multiple of the tokens (so cap it and count it), and a delegation brief needs objective, format, tools and boundaries (so a delegation tool needs a structured input, not a free string).
Changelog
| Date | Change |
|---|---|
| 2026-10-25 | Initial dossier created. |