Feature Request: Support AGENTS.md

Original Article ↗ Hacker News Discussion ↗

Anthropic’s refusal to have its Claude Code tool natively honor the emerging AGENTS.md project standard, instead insisting on its own CLAUDE.md file, has become a flashpoint for developers trying to standardize AI agent configuration across tools. Many see the technical workaround (symlinks or includes) as trivial but interpret Anthropic’s stance—and related moves like opaque prompt changes, subscription lock-in, and new bash-centric defaults—as signals of ecosystem hostility and early-stage “lock‑in” behavior. Others argue the issue is overblown given easy mitigations, but several say it has already pushed them toward alternative, more interoperable coding harnesses and open-weight models.

Scope of the Dispute

  • Thread centers on Claude Code refusing to natively treat AGENTS.md as the primary instruction file, insisting on CLAUDE.md.
  • Many see AGENTS.md as a de facto cross‑tool standard; Claude Code is viewed as the main holdout.

Workarounds & Technical Details

  • Common workarounds:
    • Symlink CLAUDE.md → AGENTS.md.
    • Put @AGENTS.md or “read AGENTS.md” as the first line of CLAUDE.md.
    • Place CLAUDE.md in .claude/ or other locations and reference AGENTS.md.
  • Some report this works fine; others say indirection (“read this other file”) leads to models intermittently ignoring instructions.
  • Claude Code auto‑loads CLAUDE.md (including nested ones) at session start, but not AGENTS.md, which only gets read after explicit prompts or includes.
  • Docs explicitly mention @AGENTS.md support, but users argue that is not the same as first‑class AGENTS.md.

Standardization vs Branding / Lock‑in

  • Pro‑standard view:
    • Other top harnesses auto‑read AGENTS.md; consistency reduces friction in multi‑agent, multi‑vendor setups.
    • Refusal is seen as petty, anti‑ecosystem behavior and a form of lock‑in or branding (“Sent from my iPhone” equivalent).
  • Counterview:
    • It’s “just a filename”; trivial to symlink or copy.
    • Complaints are overblown relative to the effort required.

Usefulness of AGENTS.md Itself

  • Supporters: file meaningfully improves workflow by encoding project‑specific conventions, gotchas, and multi‑agent usage.
  • Skeptics: concept is overhyped; cited research suggests many such context files don’t improve coding performance, especially when used as architectural dumps.
  • Some argue this info should live in human‑oriented docs (README.md, CONTRIBUTING.md); others say agents need more verbose, agent‑specific instructions than humans.

Trust, Governance, and “Developer‑Hostile” Behavior

  • Several comments describe Anthropic/Claude Code as increasingly hostile to developers:
    • Subscription discounts restricted to the vendor’s own harness.
    • Silent system‑prompt changes (e.g., pushing Bash over dedicated tools) that break workflows and aren’t clearly documented.
    • Perceived pattern of ignoring user feedback and ecosystem norms.
  • Closing the GitHub issue as “completed” without native AGENTS.md support, and odd GitHub event behavior, is seen as dishonest or evasive.

Alternatives & Exit Behavior

  • Multiple users mention moving to or preferring other harnesses: Codex, Pi, OpenCode, OhMyPi, T3 Code, custom minimal CLIs, and open‑weight models.
  • General sentiment: ecosystem is immature; users expect to switch tools often and are wary of depending too heavily on any single closed harness.