Convert to Fan-out

A monolithic agent pays frontier rates for every token, but most of an agent's tokens are mechanical — reading files, fetching pages, running tools, transcribing output. Only a thin slice is planning and judgment. The conversion moves the mechanical tokens to a cheaper model while the frontier model keeps the judgment. Anthropic's measured results:

• Advisor (Sonnet 5 executor + Fable 5 advisor): ~92% of Fable 5's SWE-bench Pro score at ~63% of the price; the advisor fires roughly once per task.
• Orchestrator (Fable 5 coordinator + Sonnet 5 workers): 96% of Fable 5's BrowseComp performance at 46% of the price; in the cookbook run the team was ~2.5x cheaper and ~3x faster than solo Fable, with 84–98% of input tokens billed at the worker rate.

Step 1 — Diagnose the current setup

Read the existing agent code or config and answer three questions:

1. Which model runs the main loop, and what share of its tokens are mechanical? Grep for the model field in API calls, agent framework config, or Claude Code settings. If a frontier model (Fable 5, Opus) is generating tool calls for bulk reading/fetching/transcribing, those tokens are the waste.
2. Is the work fan-out-shaped or long-horizon-shaped? Fan-out: independent parallel legs (research N sources, sweep M files, review K dimensions). Long-horizon: one sequential thread where each step depends on the last (a coding agent, computer use, a multi-step pipeline).
3. Where does judgment actually live? Decomposing the task, choosing an approach, verifying and synthesizing results. That's what stays on the frontier model.

Done when you can state: current model(s), the mechanical-token share, and the task shape.

Step 2 — Pick the target pattern