advisor-executorlisted

# Advisor & orchestrator patterns Two ways to get near-frontier agent quality while billing most tokens at a cheaper model's rate. Both are first-class API features, not prompt tricks. | | **Advisor** | **Orchestrator** | |---|---|---| | Shape | Cheap executor calls up to a strong advisor | Strong coordinator delegates down to cheap workers | | Feature | `advisor_20260301` tool on `/v1/messages` | Managed Agents `multiagent: {type: "coordinator"}` | | Strong model does | One-shot planning / course-correction | Planning, delegation, synthesis | | Cheap model does | All tool calls and the final output | All token-heavy reading and doing | | Anthropic's reported result | SWE-bench Pro: Sonnet 5 + Fable 5 advisor ≈ 92% of Fable-5-solo score at ≈ 63% of the price | BrowseComp: Fable 5 coordinator + Sonnet 5 workers ≈ 96% of Fable-5-solo performance at ≈ 46% of the price; cookbook measured 2.5× cheaper and 3× faster than a solo frontier agent | ## Which one Ask what the expensive part of the workload is. **Expensive part is the thinking, cheap part is the doing → advisor.** The task is one coherent line of work (fix this bug, drive this browser, run this pipeline) where most turns are mechanical but a wrong plan is costly. One process, one context, one `/v1/messages` request. Reach for this first — it is a single field in `tools` and works with the plain Messages API. **Expensive part is the reading, and it fans out → orchestrator.** The task splits into independent subtasks