opus-workers — the orchestrator reviews, cheaper tiers do the bulk work

APPLIES WHEN you are about to spawn subagents (Agent tool) or a workflow (Workflow tool) for non-trivial work AND a tier strictly cheaper than the orchestrating model can do that work faithfully.

SKIP WHEN no cheaper tier fits: the task is a single trivial step (do it inline — the spawn + review round-trip costs more than it saves), the work is a GATED action (below), or the session model is already at/below the tier the task needs.

NEVER DELEGATE gated actions. Anything bound to the user's explicit authorization — git commit, git push, publishing / public-visibility changes, deletions, account or config changes — is done INLINE by the orchestrator, never handed to a worker on any tier. The authorization lives in THIS conversation's context; a worker can't know what the user approved, and a misfire is exactly the class of action that isn't cheaply undone. (They're also single commands — spawn overhead alone exceeds the work.)

WHY. Keep the expensive orchestrating model as a THIN reviewer and move the BULK generation onto the cheapest tier that can do it faithfully. Output tokens cost more than input and re-enter context on every later turn, so offloading generation compounds across the session. A spawned reviewer call is never the cheaper option, even at a higher tier/effort than the workers: it re-bills the worker's entire output as its own input, adds its own system-prompt overhead, and buys no catch-rate edge on a rubric that's already fully specified — that's exactly why review stays folded into the orchestrator instead.

The loop

0. Pick the worker tier + effort by task class.
   • Substantive generation (real reasoning: drafting, analysis, non-trivial code): one tier down from the orchestrator (Fable → Opus; Opus → Sonnet), effort medium (routine) or high (substantive). Only when unusually demanding — correctness-critical, or you'd push toward xhigh/max — ask the user which effort before spawning.
   • Mechanical bulk (high volume, low ambiguity, judgment-free: applying an approved mechanical transform across many files, genericizing copies, regenerating fixtures, mass renames per an approved plan): two tiers down (Fable → Sonnet; Opus → Haiku), effort low or medium. If a "mechanical" task turns out to need judgment mid-flight, that's a misclassification — pull it back up a tier or inline; don't let the cheap worker improvise.
   • Rule of thumb: the cheapest tier the task class tolerates, always strictly cheaper than the orchestrator; if nothing cheaper fits, do it inline.
   • Effort binds on the Workflow path (agent(prompt, { model, effort })); the plain Agent tool has no effort param (subagents inherit session effort) — note that in your summary when it applies.
1. Pre-register the review rubric — BEFORE spawning. Write down, in your plan or the spawn message, the 3–7 concrete checks this output must pass and HOW each will be stress-tested: the command you'll run, the canary, the sample you'll re-derive by hand, the diff you'll scan. Executable checks over eyeball checks wherever possible. The same rubric goes into the worker prompt as its success criteria. Pre-registering kills the two review failure modes: rubber-stamping, and a post-hoc rubric quietly shaped by whatever the output happens to be. Scale it to the tier: substantive work gets the full rubric; mechanical bulk gets counts + spot-checks (N samples re-derived, diff scanned for out-of-scope lines) — a line-by-line re-read of bulk output eats the savings.
2. Delegate. Spawn every worker with the step-0 model override — never the orchestrator's model. Each worker gets a self-contained task AND the step-1 rubric as its success criteria. Independent tasks → fan out in parallel.
3. Review as the orchestrator, against the pre-registered rubric. Run the rubric as written — every check actually executed. You may ADD checks for problems you only saw once the output existed; you may NOT silently drop a pre-registered check that is failing. The orchestrator IS the reviewer — don't spawn one and don't downgrade the review.
4. Decide per output.
   • Satisfactory (rubric passes) → keep it. Once all workers pass, present.
   • Not satisfactory → send back to the SAME worker with the failing rubric items and the concrete change you want. "Do better" wastes a round — name the fix.
5. Bound the loop. At most 2 redo rounds per worker. Still short after that → present the best version WITH an honest note on what's still weak. Never loop forever; never silently ship a result you know is weak.

Rules

• The routing is an instruction you follow, not something the harness enforces — you must actually pass the model override when you spawn. Forget it and the worker runs on the orchestrator's model and the savings are gone.
• Review is the orchestrator's job precisely because it's the quality gate. Delegating the review itself to a cheaper tier defeats the purpose.
• The rubric is written before the output exists. A review that invents its criteria after seeing the output isn't a gate, it's a rationalization.
• Worth it only when the delegated work is substantial (multi-file, multi-step, or fanned out). For a quick lookup or a one-line edit, skip the whole dance.
• Savings come from the workers doing the output-heavy generation; the orchestrator still pays input tokens to read each output during review. The win scales with how much the workers GENERATE vs. how much the orchestrator re-reads — which is why bulk review is sampled, not re-read line-by-line.
• All standing rules (no fabrication, surgical scope, project conventions) apply to every worker tier and the review — delegation doesn't relax them.