diff --git a/agents/AGENTS-claude.md b/agents/AGENTS-claude.md index b2a48b3..2be0f12 100644 --- a/agents/AGENTS-claude.md +++ b/agents/AGENTS-claude.md @@ -41,21 +41,27 @@ Prefix every `Agent` prompt with one of: - `[TIER: STANDARD]` — focused implementation logic; minimal preamble - `[TIER: DEEP]` — complex reasoning; must conclude with a summary -### Model selection (override only when needed) +### Model selection (relative to the parent, no fixed names) Resolution order: explicit `model` param → agent type frontmatter → parent inheritance. (`fork` always runs on the parent model; an override is ignored.) +Model names shift (Opus was the top tier when this was written; Fable and +Mythos sit above it now), so think in tiers, not names: 0. **Self-hosted backend (a SessionStart hook says so when `ANTHROPIC_BASE_URL` - points at the vLLM gateway): never pass `model`.** Aliases like `sonnet` - do not exist there; every subagent inherits the one available model. -1. **Don't override specialized agent types** (`Explore`, `Plan`, etc.) — their frontmatter is tuned. Leave `model` off. -2. **For `general-purpose` under an expensive parent (Opus/Fable): set `model: sonnet`** — don't pay top rates for routine subagent work. -3. **`opus`** for research, recommendation, trade-off and judgement work - (also blinded eval judges) and when isolated context *and* reasoning beyond - the parent's inline capacity are both needed. For "how should X look" - questions discuss options first, do not build. -4. **`model: haiku`** only for mechanical tasks: bulk classification, format conversion, summarizing pre-filtered text. Code-semantic work → Sonnet. + points at the vLLM gateway): never pass `model`.** Aliases do not exist + there; every subagent inherits the one available model. +1. **Specialized agent types** (`Explore`, `Plan`, …) keep their frontmatter + model. Leave `model` off. +2. **Routine `general-purpose` work** (mechanical edits, bulk lookups, + verification runs): one tier below the parent, today `sonnet`. Don't pay + top rates for routine work. +3. **Judgement work** (research, recommendation, trade-off, review, blinded + eval judges): the parent's tier, so pass no `model` and let it inherit. + For "how should X look" questions discuss options first, do not build. +4. **`haiku`** only for purely mechanical tasks: bulk classification, format + conversion, summarizing pre-filtered text. Code-semantic work is at least + `sonnet`. ### Context hygiene