---
name: model-tiering
description: Assign the right model tier to each job — expensive brains plan and judge, cheap hands search and grind. Use when dispatching any sub-agent, configuring a multi-agent workflow, or when quota pressure appears.
---

Brain work and muscle work are different jobs. Paying brain prices for muscle work is how a week's quota disappears in days.

## The split

- **Brain (top model):** planning, final synthesis, hard-bug diagnosis, merge-conflict resolution, the deciding vote in a review.
- **Muscle (cheapest model that survives the task):** search, fetching and summarizing, per-item verification, formatting, translation, running tests.
- **Middle tier:** routine implementation against a clear spec — a mid model with tests and review beats a top model without them.

## Rules

- **Sub-agents inherit the parent's model unless told otherwise.** The parent is usually your most expensive session — *always set the model explicitly* on every dispatch. Forgetting this once cost us a 20-agent research fleet all running on the flagship.
- **The three-call rule.** In a long-running expensive session, any chore that will take more than ~3 tool calls (inventory sweeps, batch edits, running test suites, install-and-verify) goes to a cheap agent as one package — even when each call looks tiny. A fat session re-reads its whole history on every turn, so in a long chat there is no such thing as a small chore. (We learned this the same day we wrote the meters rule: the meter drifted anyway, and the leak was "quick" hand-done chores in a day-old session.) Don't trust willpower here — this repo ships a hook that enforces it mechanically: see [hooks/](../../hooks/README.md). No self-granted exceptions: the only inline acts are conversation, the final artifact's single Write, one verification command, and reading agent results — everything else dispatches.
- **Scouting is muscle work.** Reading files or web pages to summarize them never needs the top model. The expensive session reads conclusions, not raw material.
- **Escalate the hard case, not the batch.** If one item in a cheap batch turns out to be genuinely hard, send that one item up a tier. Don't promote the whole pipeline.
- **Spend to the meters, not to habit.** "Cheap hands" is a pressure response, not a religion. Once dispatch discipline has cut your fleet frequency down, check the meters: with comfortable headroom at your actual cadence, default your workers to the strongest tier the budget sustains — you are buying fewer errors and fewer retries, which is usually cheaper than redoing sonnet-grade work. Downshift when the meters tighten. The meters set the tier; frugality theater does not.
- **Keep the meters marching together.** If your plan shows per-tier usage meters, read them at every dispatch: top-tier meter running ahead → your main session is doing muscle work itself, push drafting/digging down a tier; aggregate meter running ahead → your fleets are too fat, shrink them and do small jobs inline. No meter should hit 100% while another sits half full.
- **Quota pressure shrinks batches, not thinking.** Under a tight budget, cut fan-out and skip nice-to-have verification passes — never dumb down the planning or the final judgment.

## Success criteria

Every dispatch names its model on purpose. The bill's biggest line is thinking, not grinding. No flagship tokens were spent reading raw files.