User directive on how to run large subagent fan-outs (deep-research workflows, verification panels, any bulk agent fleet), captured 2026-07-03 during the slv1 Strategy #1 research.
Directive
Run subagent fleets on Opus, never Fable, for bulk fan-outs. Fable burns through usage limits fast on large fleets. Opus is the default tier for fan-out work. (Complements the standing global rule: never Sonnet for any subagent.)
Slimmed Workflow Shape
Prefer the slimmed deep-research shape over the maximal one — it delivers comparable decision value at roughly half the tokens.
| Shape | Angles | Sources | Verification | ≈ Agents | ≈ Tokens |
|---|---|---|---|---|---|
| Slimmed (preferred) | 4 | 10 | 12 claims × 2 votes | ~56 | ~1.75M |
| Maximal | — | 21 | top 25 × 3 votes | ~103 | ~3.24M |
Weaker-quorum tradeoff
The 2-vote panel is a weaker quorum than 3-vote: a 1-1 split survives as “confirmed.” Treat single-vendor findings from a slimmed run as existence proofs, not accuracy claims. Reach for the maximal shape only when a claim must be load-bearing for a large capital or infra commitment.
Provenance
Both shapes were run back-to-back on the same question (slv1 copy-trading feasibility). The slimmed round-2 run closed most gaps left by the maximal round-1 run with zero agent errors, at ~54% of the token cost. See slv1-copy-trading-feasibility-2026-07-03 for the substantive output and how the two rounds combined.
Related
- slv1-copy-trading-feasibility-2026-07-03 — the research that produced this directive
- obsidian-subagent-improvements-2026-05-04 — related subagent-system conventions