From 7a0fcf855e9256b85bbc8776b190d8bc78501ba1 Mon Sep 17 00:00:00 2001 From: Andy Bonventre Date: Wed, 12 Aug 2026 16:31:32 -0400 Subject: [PATCH] docs(agents): state the subagent model split as capability tiers The build-vs-review split is a tier rule, not a model-name rule: fast mid-tier for implementation, strongest tier for review and judgment. Claude Sonnet/Opus is today's mapping; re-map when the family changes. --- AGENTS.md | 33 ++++++++++++++++++++------------- 1 file changed, 20 insertions(+), 13 deletions(-) diff --git a/AGENTS.md b/AGENTS.md index d7e2a38..ad8b537 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -1363,19 +1363,26 @@ Rules: ## Subagent model strategy When you spawn subagents, set the model EXPLICITLY on every spawn — never -let an implementation agent inherit the parent model by default. - -- **Code-writing / implementation / mechanical work uses Sonnet.** Writing - code, editing docs, changing config, grep/investigation, watching a - deploy — spawn these with the Sonnet model. Sonnet is fast and sufficient. -- **Review, adversarial verification, and judgment gates use Opus.** The - review-to-zero gate and any correctness verdict deserve the stronger - model. Spawn review agents with Opus. - -The default pattern for a change: a Sonnet agent writes the PR, an Opus -reviewer drives it to zero. Omitting the model makes a subagent inherit the -parent's model, so an Opus parent silently runs implementation work on Opus -— expensive and backwards. Pass the model on every spawn. +let an implementation agent inherit the parent model by default. The rule +is a capability-tier split, not a vendor or model-name rule; it applies to +whatever frontier family is current. + +- **Code-writing / implementation / mechanical work uses the fast + mid-tier.** Writing code, editing docs, changing config, + grep/investigation, watching a deploy — the tier that is fast, cheap, + and sufficient for well-specified work (today: Claude Sonnet; the + equivalent tier elsewhere: GPT mini/frontier-fast class, Gemini Flash). +- **Review, adversarial verification, and judgment gates use the + strongest tier.** The review-to-zero gate and any correctness verdict + deserve the strongest available model (today: Claude Opus; elsewhere: + the full frontier flagship, never a mini/fast variant). + +The default pattern for a change: a mid-tier agent writes the PR, a +strongest-tier reviewer drives it to zero. Omitting the model makes a +subagent inherit the parent's model, so a strong parent silently runs +implementation work at flagship price — expensive and backwards. Pass the +model on every spawn. When a model family changes, re-map the two tiers +and keep the split; do not carry a stale model name forward. ## One agent per plan task, not one agent per plan