Claude Code now reads AGENTS.md if there is no Claude.md

Original Article ↗ Hacker News Discussion ↗

Claude Code has added support for reading a project’s AGENTS.md file when no CLAUDE.md is present, aligning it with an emerging de facto standard used by other AI coding tools. Developers welcome the change mainly because it reduces lock‑in and the need for symlinks or duplicated config files, but many criticize Anthropic for taking so long and still ignoring related conventions like `.agents/skills` and user‑level agent configs. The exchange reflects broader frustration with proprietary harness behaviors, shifting market power among AI vendors, and the slow, messy emergence of shared standards for agent configuration.

Scope of the change

  • Claude Code now falls back to AGENTS.md when no CLAUDE.md is present.
  • Many commenters say this lets them delete one-line CLAUDE.md stubs or symlinks that just pointed to AGENTS.md.

Standards vs proprietary behavior

  • AGENTS.md is described as a de facto standard across many harnesses; Claude’s insistence on CLAUDE.md was seen as pushing a proprietary workflow and branding.
  • Some argue it’s “about time” Claude respected the shared convention; others note CLAUDE.md actually predated AGENTS.md but was quickly outvoted by ecosystem adoption.
  • A minority argues prompts should be tuned per model and that a universal AGENTS.md may hurt model diversity.

Remaining gaps and workarounds

  • Many emphasize this is only a partial fix: Claude Code still doesn’t support .agents/skills (or ~/.agents/skills), which most other harnesses do.
  • Users share workarounds: symlinking .agents/skills into .claude/skills, Git hooks to automate that, and using CLAUDE.md as @AGENTS.md plus Claude-specific additions.
  • Windows/git symlink limitations complicate some of these approaches.

Instruction-following reliability

  • Several note that even with CLAUDE.md/AGENTS.md, Claude often ignores instructions (e.g., CLI flags, “don’t add ‘Made with Claude’ text”).
  • Comments point to the harness system prompt that labels these files as “optional context,” which may reduce compliance compared to other tools that treat AGENTS.md as mandatory.

Vendor lock‑in and harness choice

  • There is strong skepticism toward closed-source, provider-controlled harnesses and their incentives to maximize token usage and lock users in.
  • Some report switching to alternative harnesses and/or open-weight models, citing better adherence to AGENTS.md and .agents/skills, portability across providers, and more user-aligned design.

Meta reactions

  • Many see this as the bare minimum rather than something to celebrate, but still welcome the move toward a shared standard.
  • The intensity of discussion over a small feature is taken as a sign of how central AI harness ergonomics and interoperability have become.