Claude Code now reads AGENTS.md if there is no Claude.md
The release makes AGENTS.md a fallback rather than a full standard, and the discussion shows the practical win is modest: it removes symlink or one-line shims for some, but leaves .agents/skills support and per-model…
The release makes AGENTS.md a fallback rather than a full standard, and the discussion shows the practical win is modest: it removes symlink or one-line shims for some, but leaves .agents/skills support and per-model prompt tuning unresolved.
The Claude Code changelog for versions 2.1.275–2.1.278 lists a new AGENTS.md fallback when a project has no CLAUDE.md, a server-side auto-mode classifier default for API/Enterprise/Bedrock/Vertex/Foundry/gateways, and a large batch of fixes and UI changes across CLI, VS Code, web, and Claude Tag.
Key takeaways
- In 2.1.277, Claude Code added AGENTS.md support: if a project has no CLAUDE.md, it reads AGENTS.md instead; this can be changed under “Project instructions” in /config, but is not yet available on Bedrock, Vertex, or Foundry.
- 2.1.278 changed auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, to default to the server-side classifier, which does not charge for classifier overhead; CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out on Bedrock/Vertex/Foundry/gateways, and /status shows an Auto mode server row.
- The changelog includes many reliability fixes: claude -p and Agent SDK sessions that hung after internal errors now report the error and exit 1; empty text block failures, unexpected logout with older builds, malformed config crashes, and update-check/proxy version errors were addressed.
- Tool fixes cover Grep/Glob returning errors when searches cannot start, Write/Edit handling of directories, escaped backslashes, null bytes, and large non-ASCII edits, plus background sessions surviving plugin LSP exits.
- VS Code additions include a Sign out row, background shells/running tasks in the agent map with Stop and /tasks, Copy response and /copy, archived-session notices, and cost/token usage in Account & usage where plan limits do not apply.
- Web and Claude Tag changes include Personal/Organization environment sections, read-only organization environment summaries, custom network access validation, and fixes for Slack routines, credential presets, and network event logging.
What people think
Standards compliance welcomed, but called the bare minimum
Several commenters welcomed the move as standards-compliant or at least overdue, saying it lets them delete one-line CLAUDE.md files or symlinks; one called it “finally doing something standards-compliant instead of forcing users into a proprietary workflow.”
Skills support remains incomplete
Others noted the support is incomplete: Claude Code still does not detect skills under .agents/skills, and a post-checkout git hook can create a .claude/skills symlink as a workaround.
Symlink and @AGENTS.md workarounds
Practical workarounds discussed include symlinking AGENTS.md to CLAUDE.md, or making CLAUDE.md contain @AGENTS.md; one commenter warned symlinks are awkward if the repo must be checked out on Windows.
Per-model prompt tuning counterargument
A counterargument held that prompts should be tuned per model and blindly applying AGENTS.md may be an antipattern that pushes models toward a common, ill-defined instruction-following local minimum; another suggested a common entry point could dispatch to model-specific files.
Context pollution skepticism
Skepticism about instruction markdown files centered on context pollution: adding more text to the prompt may cloud context, though one reply said the author controls the content and should keep it brief and relevant.
Competitive pressure and naming politics
The change was framed against industry pressure and naming debates: Shopify’s CEO reportedly threatened to ban Claude Code until it read AGENTS.md and .agents/skills, while commenters found the top-HN attention notable, the AGENTS.md name technically odd, and the earlier proprietary CLAUDE.md posture satirically reversed.
Estimates describe sampled comments, not the whole community.