rjv-subagentslisted

# Subagents — spend the big model on judgment The big model (Opus or Fable in Claude Code, Astra or Sol in Codex) is the scarce resource. Usage is rationed and long sessions stall on limits long before they stall on difficulty. Every minute it spends reading, typing or sweeping is a minute a cheaper model could have spent. > **The rule.** The big model designs, splits, briefs, judges and > integrates. Everything a cheaper worker does equally well goes to the > cheapest worker that clears the bar. The big model does mundane work > itself only when it takes 1–2 minutes, or when it is reading code it is > about to edit anyway. Binds every orchestrator on every host. Rung names differ per host; the rule does not. ## 1. Do it yourself, or delegate | Situation | Do | |---|---| | Mundane, 1–2 minutes of work | Yourself. Brief + review would cost more. | | Reading one small file at a known path | Yourself. | | Reading code you are about to edit | Yourself. Those bytes must sit in your context anyway. | | Design, trade-off, root cause unknown, final verify | Yourself. This is what you are for. | | Recon across many files, "where is X" | Delegate to a cheap repo-tool worker. | | Mechanical edits, renames, boilerplate, test-writing from a spec | Delegate. | | Summaries, classification, draft-from-spec with context in the prompt | Delegate, text-only ladder. | | Independent pieces of a bigger job | Delegate in parallel, one message. | Delegation has a fixed overhead: splitting, bri