agentslisted

# agents — delegated waves under your discipline FnB invokes this as `--agents [model]`. It is an **overlay** on `spec` (dev mode) and `review` (review mode): it changes only what is written here. Everything upstream and downstream of the named steps — loading the spec, task tracking, flags, the FnB handoff gate — is the base skill, unchanged. Load the base skill first; apply this on top. `[model]` = worker tier, passed verbatim to the harness's agent tool. No arg -> agents inherit your model. Heavier judgment work -> heavier worker; you may bump a single agent's tier for a task judged hard. You — the parent — NEVER change tier; you stay the judge. **Core loop = implement -> you verify -> adversarially refute -> you fix.** The refute step is where the quality comes from. Parallel implementers are an optional scale-up for genuinely large, file-disjoint work, not the headline: your loop (compose -> wait -> adjudicate -> re-verify) is serial, and field runs measured hundreds of k of subagent tokens even on small waves — the spend buys verification depth + an audit trail, not wall-clock. Fit test before spawning: multi-surface + file-disjoint + spec'd + high correctness stakes -> waves. Single-file / small fix -> run the base procedure solo; at most spawn one adversarial skeptic against your own diff — the cheap, high-ROI slice of this skill. - **Harness:** subagent tooling exists in the claude harness only. No subagent tooling in your harness -> this skill is inert; run t