brain/
conceptartificial-intelligence

AGENTS.md vs CLAUDE.md

Notes

Vintage: 2026-08. Instruction-file standard fight as of 25 Aug 2026 in 2026-08-27-best-agent-harnesses-for-programming. Claude Code’s promised AGENTS.md support was “working on it,” not fetched as shipped.

AGENTS.md vs CLAUDE.md

One-line summary: Two competing instruction-file standards for programming agents — OpenAI documents AGENTS.md as the instruction chain; Anthropic still ships CLAUDE.md as native. The 25 Aug 2026 X fight is primary (Tobi vs Thariq), not a settled industry standard.

The insight

This is a standard question, not the same as “should a team keep a living CLAUDE.md?” (claude-md-team-knowledge-base). The process claim (update the file when the agent errs) can survive either filename. The contested claim is which filename a cross-tool repo should commit, and whether one model family can read the other’s file without a translation layer.

Evidence

  • From 2026-08-27-best-agent-harnesses-for-programming: Codex official AGENTS.md is the instruction chain — global ~/.codex, then repo-root-to-cwd walk, override files, 32 KiB default cap, fallback filenames (project_doc_fallback_filenames).
  • From the same source: Claude Code official is CLAUDE.md / .claude/ skills; dynamic workflows mine sessions back into CLAUDE.md.
  • From the same source (X, primary, 2026-08-25): Shopify CEO Tobi Lütke threatened to ban Claude Code until it reads AGENTS.md.
  • From the same source (X, primary, 2026-08-25): Claude Code staff Thariq Shihipar said they will add AGENTS.md / system-prompt hackability, but argued model families are not interchangeable so Claude.md stays default.
  • From the same source (X, discourse, 2026-08-25): Kun Chen — “industry already has a standard.” Discourse, not a lab measurement.
  • From the same source: Claude Code’s promised AGENTS.md support was “working on it” as of 25 Aug 2026, not fetched as shipped.

This thread already had a weaker AGENTS.md signal: claude-md-team-knowledge-base notes a reply-commenter variant (AGENTS.md for agent behaviors) and Cursor Cloud Agent best-practices using agents.md + skills. Those are process/filename color, not the official Codex instruction-chain spec.

Contradictions / tensions

  • Official standard vs native default. OpenAI documents AGENTS.md as the instruction standard. Anthropic still ships CLAUDE.md as native. Both can be true at once; neither source fetched a shipped Claude Code AGENTS.md reader.
  • Interchangeability. Thariq’s load-bearing argument is that model families are not interchangeable, so a shared filename would still need a Claude-shaped body. Tobi’s load-bearing demand is that Claude Code read the file the rest of the industry already writes. Unresolved.
  • This thread’s own repo. Brain already carries both AGENTS.md (Grok) and Claude.md (Claude Code compat). That is local practice, not evidence that the industry fight is settled. Do not treat vault conventions as a source.

Open questions

  • Did Claude Code ship AGENTS.md / fallback-filename support after 25 Aug 2026?
  • If it ships, does Claude treat AGENTS.md as a first-class walk or as an alias that still prefers CLAUDE.md?
  • Does a shared filename without a shared body actually port instructions across model families?

Sources

Related

Referenced by