Skip to content

docs(agents): state the subagent model split as capability tiers - #132

Merged
andybons merged 1 commit into
mainfrom
andybons/agents-model-tiers
Aug 12, 2026
Merged

docs(agents): state the subagent model split as capability tiers#132
andybons merged 1 commit into
mainfrom
andybons/agents-model-tiers

Conversation

@andybons

Copy link
Copy Markdown
Contributor

Generalizes the Sonnet/Opus rule across frontier model families: implementation uses the fast mid-tier, review/judgment uses the strongest tier, with today's Claude mapping named as the current instantiation and an explicit re-map-on-family-change instruction so the rule survives model generations and provider swaps.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WG6N8YSSwXMma1DKTMwJdf

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.
@claude

claude Bot commented Aug 12, 2026

Copy link
Copy Markdown

Claude finished @andybons's task in 52s —— View job


Review

  • Read AGENTS.md diff and surrounding context
  • Review for correctness, consistency, and STE writing style
  • Post review

Verdict: LGTM. Docs-only edit to AGENTS.md. The generalization from a hard Sonnet/Opus rule to a capability-tier split is faithful and internally consistent.

What holds up well:

  • The core invariant is preserved: set the model explicitly on every spawn, implementation on the fast mid-tier, review/judgment on the strongest tier. The "omitting the model makes a subagent inherit the parent" warning survives intact, just reworded from "on Opus" to "at flagship price."
  • The added re-map instruction — "When a model family changes, re-map the two tiers and keep the split; do not carry a stale model name forward" — is the right durability hook and matches the file's own "a behavior change updates AGENTS.md in the same commit" ethos.
  • Today's Claude mapping (Sonnet ↔ Opus) is retained parenthetically, so nothing is lost for the current instantiation.
  • No other Sonnet/Opus references remain in AGENTS.md (grep confirmed lines 1363–1385 are the only ones), so there's no stale contradiction left behind.

Minor, non-blocking observations (style only — take or leave):

  • The strong tier is named three slightly different ways in three sentences: "the strongest tier", "the strongest available model", "the full frontier flagship", plus "flagship price". AGENTS.md's own STE rule is "one word for one idea — pick a term and reuse it verbatim." Settling on one label (e.g. always "the strongest tier") would tighten it, but the meaning is unambiguous as written.
  • GPT mini/frontier-fast class, Gemini Flash are illustrative cross-vendor examples; they read fine as examples and the "not a vendor or model-name rule" framing correctly signals they aren't normative.

No correctness, concurrency, or protocol concerns — this touches no code. Nothing to fix.
andybons/agents-model-tiers

@andybons
andybons merged commit f57786b into main Aug 12, 2026
2 checks passed
@andybons
andybons deleted the andybons/agents-model-tiers branch August 12, 2026 20:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant