Buried in the September 18 Claude Code changelog, between gateway header fixes and a stale .git/config.lock bug fix, is one line that matters more than its length suggests:
“Added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead; change it under ‘Project instructions’ in
/config(not yet on Bedrock, Vertex or Foundry)”
That’s v2.1.277. It’s not a redesign, it’s a fallback: if a repository has no CLAUDE.md, Claude Code will now read AGENTS.md instead of nothing. CLAUDE.md still wins whenever it exists. It’s not yet available on Bedrock, Vertex, or Foundry deployments. On paper, that’s the whole feature.
It’s still worth a full article, because of what AGENTS.md is, who’s behind it, and how long Anthropic sat on a change this small.
What AGENTS.md actually is#
AGENTS.md is a convention, not a product — a plain-markdown “README for agents” that a coding agent reads at session start to learn a project’s build commands, testing conventions, and dos-and-don’ts. It was published in August 2025 out of parallel work by the teams behind OpenAI Codex, Sourcegraph’s Amp, Google’s Jules, Cursor, and Factory — explicitly framed on its own site, agents.md, as an attempt to avoid “five competing names for the same file.” It worked, mostly: the spec’s own count now lists more than 60,000 open-source projects and 25-plus supporting tools, including OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Google Jules, Devin, Aider, Zed, Windsurf, and JetBrains Junie.
On December 9, 2025, OpenAI donated AGENTS.md to the Agentic AI Foundation — the same Linux Foundation body that now stewards MCP, which this blog covered when Anthropic donated MCP to it in May. Here’s the detail worth sitting with: Anthropic is a listed co-founder of the AAIF, alongside Block, with Google, Microsoft, AWS, Bloomberg, and Cloudflare as supporting members. Anthropic has had a governance seat over the exact convention it just started reading — for over nine months — before Claude Code actually parsed a single AGENTS.md file.
That’s not a scandal. It’s a genuinely small engineering task that apparently sat behind higher-priority work, which is a defensible call for any team shipping as fast as Anthropic’s Claude Code group does. But it’s also a useful corrective to the version of this blog’s own argument that sometimes runs too clean: co-founding a standards body doesn’t automatically translate into fast product follow-through, even for the vendor most invested in agentic tooling being taken seriously as infrastructure.
Why this isn’t just a bookkeeping change#
For any team running more than one coding agent against the same repository — a common pattern now that Cursor, Copilot, Codex, and Claude Code all show up on the same engineering org’s tool list — AGENTS.md fallback support means one less place for instructions to silently diverge. Previously, a repo maintaining both AGENTS.md (for Codex, Copilot, Cursor users) and CLAUDE.md (for Claude Code users) had two files to keep in sync, with no guarantee anyone remembered to update both. Now a team that hasn’t written a CLAUDE.md at all gets Claude Code reading the same instructions everything else already reads, for free.
The reaction on Hacker News split along predictable lines: some readers called it “the absolute bare minimum” and asked what took so long, others welcomed it as Anthropic “finally doing something standards-compliant.” Both are fair. It is the bare minimum, and it is also a genuine practical win for anyone who’s had to explain to a new teammate why their AGENTS.md-following agent behaves differently from the one that only reads CLAUDE.md.
The part the changelog line doesn’t mention#
This blog spent the better part of April and a dedicated guide in May making one argument about CLAUDE.md: it is not documentation, it is an auto-loaded, trusted-by-default instruction source, and CVE-2026-21852 proved that treating it otherwise is how you get a supply-chain attack disguised as a project config file. The fix at the time was narrow — Anthropic patched the specific 50-subcommand deny-rule parsing bug — but the broader lesson this blog kept returning to was structural: any file an agent reads and acts on without a human re-reading it first is executable, whatever its extension says.
AGENTS.md is that same trust boundary, extended across a much wider surface. A poisoned CLAUDE.md is a Claude Code problem. A poisoned AGENTS.md — sitting in a repo that’s also read by Codex, Copilot, Cursor, Jules, Devin, Aider, and 20-odd other tools listed on agents.md’s own supporter page — is a cross-vendor problem the moment any one of those agents checks it out. Nothing in Anthropic’s changelog entry, nor anything found in Claude Code’s public documentation as of this writing, suggests AGENTS.md gets different scrutiny than CLAUDE.md does at read time. It inherits the same blast radius, on a file format explicitly designed to be picked up by everyone at once.
None of this means don’t ship it. Interoperability is worth the marginal risk, and the alternative — Claude Code being the one agent in a multi-agent repo that silently ignores half the project’s stated conventions — is its own kind of failure. But teams treating AGENTS.md as “just markdown” because it predates and outlives any single vendor’s security posture are making the same mistake this blog flagged with CLAUDE.md eighteen months ago, at a larger scale.
The actual takeaway#
Claude Code adopting AGENTS.md is correct, overdue, and small. It closes a real interoperability gap for teams juggling multiple agents in one repo, it’s consistent with Anthropic’s own governance commitments inside the AAIF, and the changelog entry undersells how much it will quietly help anyone who’s ever had two conflicting instruction files pointed at the same codebase. It’s also a reminder that “we co-founded the standards body” and “we shipped support for the standard” are two different milestones, separated here by the better part of a year — and that every new auto-loaded instructions file, however boring, is worth the same scrutiny this blog gave CLAUDE.md the first time around.
Sources: Claude Code changelog (primary, v2.1.277 entry, Sept 18, 2026); agents.md (canonical spec site, origin and adopter list); OpenAI — Agentic AI Foundation announcement and Linux Foundation press release (Dec 9, 2025, AAIF founding members); Hacker News discussion; this blog’s prior coverage of the CLAUDE.md Trap (CVE-2026-21852), the Complete CLAUDE.md Guide, and MCP’s donation to the Agentic AI Foundation.