TL;DR: Claude Code finally reads AGENTS.md natively as of v2.1.277. A community kept a two-line feature request alive for 13 months, and won. Here's how it happened.
On August 21, 2025, a developer named DylanLIiii filed what looked like a small feature request against Claude Code. The ask was simple: support AGENTS.md alongside CLAUDE.md.
Two days earlier, a group of toolmakers had published agents.md, a spec for a single, shared instructions file that any coding agent could read. Codex, Cursor, Amp, and others were adopting it. The request pointed out the obvious friction: CLAUDE.md works fine if everyone on a team uses Claude Code, and not at all when half the team is on something else.
The issue became the most upvoted in the repository's history. 357 thumbs-up, 405 comments, and a year-long saga that says a lot about how open standards actually win. Not through committees, and not through vendors deciding to play nice. Through a community that refused to let a small, obviously-correct request die quietly.
A year of silence — and a year of showing up
For roughly twelve months, Anthropic did not comment on the issue. Not to reject it, not to accept it, not to explain. Nothing.
Most issues die at this stage. This one didn't, because the community kept showing up, and kept building. They symlinked CLAUDE.md to AGENTS.md. They used Claude Code's @-import syntax to make a one-line CLAUDE.md that pulled in AGENTS.md. They wrote SessionStart hooks that injected the file manually. Every workaround was a small monument to the fact that the native fix was, as one commenter put it, "two lines of code." Every workaround was also a public, working demonstration that the request was reasonable.
The comments from that period read like a slow-burning case study in collective pressure done right. Nobody rage-quit or drifted off. They documented, they upvoted, they asked again:
"I hope the team won't ignore it and adopt it ASAP to show the will to make the right decisions rather than ones that benefit the business. It is a small issue, but because of that it is telling." — kossnocorp, September 2025
"It's very likely a business driven decision not to support this. CLAUDE.md serves as a network effect every time someone opens a github repo and sees it." — JoeRoddy, September 2025
That second comment is probably the best read on the delay. CLAUDE.md files scattered across thousands of repos are free advertising. Every developer who opens a project and sees CLAUDE.md at the root has just been reminded which agent their colleagues expect them to use. A generic AGENTS.md gives that up.
The community saw that dynamic clearly, and named it in public. Companies hate that.
The close that wasn't
Then, on August 17, 2026 — almost exactly a year after the issue was filed — an Anthropic maintainer finally responded. Not with a fix. With a close.
"Claude Code reads CLAUDE.md, but you can share one file with other agents: create a CLAUDE.md containing just @AGENTS.md (an import), or symlink CLAUDE.md to AGENTS.md." — bcherny, closing the issue as "completed"
The community's reaction was swift and, honestly, deserved. The response pointed to the same workaround people had been using for a year, dressed up as a resolution. It required keeping a CLAUDE.md in every directory that had an AGENTS.md, which is a real problem when you consider that OpenAI's own repository contains 88 AGENTS.md files. The fix, as one commenter noted, would have meant adding 88 dummy CLAUDE.md files.
"We expected better from you than some lame-ass workaround and a silent close." — cobrabr, August 2026
Here's the part worth noticing: the close didn't end the issue. It re-energized it. A community that had spent a year building workarounds knew exactly what a real fix looked like, and they weren't about to accept a workaround with a "completed" label on it. The issue stayed closed for about a month because people wouldn't let it stay resolved-in-name-only.
The reversal
On September 18, 2026, Anthropic shipped Claude Code v2.1.277 with native AGENTS.md support. If a project has no CLAUDE.md, Claude Code now reads AGENTS.md instead. Same locations, same subdirectory behavior, same import expansion. A /config toggle controls the behavior, including a mode that loads both files side by side.
The implementation is itself interesting: it shipped as a built-in "mod," part of a new extensibility system for Claude Code's harness, with source available. The docs now explicitly say the symlink and import workarounds are unnecessary. The file's content is read once either way.
Thirteen months after the request. One month after the close.
The part that makes it a story
Here's the detail that turned a routine feature request into a public relations problem: in late 2025, Anthropic co-founded the Agentic AI Foundation under the Linux Foundation, alongside AWS, Google, Microsoft, OpenAI, Cloudflare, and others. One of the foundation's initial contributed projects? AGENTS.md.
So for the better part of a year, Anthropic was simultaneously a founding steward of the standard and a non-adopter of it. The community noticed, and asked the question out loud:
"Why would Anthropic cofound an AI standards organization which includes AGENTS.md and not even have a basic level of support for it in their products?" — alexgleason, December 2025
The most likely explanation is commercial. CLAUDE.md was a differentiator, product priorities were elsewhere, and nobody wanted to be the person who gave up a branding surface. But the gap between the standards press release and the issue tracker was visible to everyone, and issue #6235 became the place where that was loudest.
What actually won
The @AGENTS.md import workaround meant the feature was never technically blocked. What changed was people, organized and patient and working in public:
The standard won elsewhere, decisively, because the community adopted it. Codex, Cursor, Gemini CLI, Copilot, Zed, Windsurf, Jules, Amp, and dozens of others adopted AGENTS.md. The site now claims 60,000+ projects. Every one of those adoptions was a developer or a team choosing a shared standard over a vendor-specific file. At some point, being the holdout stopped being a differentiator and started being a tax. And that tax was levied by users, not regulators.
The pressure stayed public and specific. The issue never became a vague pile-on. It stayed a concrete, small, obviously-correct request with 357 upvotes attached, backed by people who documented every workaround and every dead end. That's much harder to dismiss than general grumbling. It's also much harder to make disappear.
The close backfired, because the community wouldn't accept it. Closing an issue as "completed" while pointing at a workaround is the kind of move that converts neutral observers into critics. The people in the thread knew the difference between resolved and relabeled, and they said so. The month between the close and the shipping feature was, presumably, not a coincidence.
If you're using the workaround today
If AGENTS.md is your team's source of truth, you can drop the stub CLAUDE.md files and symlinks once everyone is on v2.1.277 or later. Claude Code now falls back to AGENTS.md in the same locations, including subdirectories.
Two thing to note. First, the fallback only kicks in when there's no CLAUDE.md, so a leftover stub will keep taking precedence. Second, anyone on an older version still needs the import or symlink. If you want Claude-specific instructions on top of the shared file, check the /config toggle for the mode that loads both side by side.
The takeaway
The lesson of issue #6235 isn't really about Anthropic, or even about AGENTS.md. It's about what a community can do when it stays organized and refuses to go away.
One developer filed a two-line feature request. Hundreds of others upvoted it, commented on it, built workarounds around it, and documented why it mattered. Nobody had formal power. Nobody got paid. There was no committee, no petition, no coordinated campaign. Just a public thread that refused to go quiet. Thirteen months later, one of the largest AI companies in the world shipped the fix, deprecated its own workarounds, and published the source.
That's the playbook, and it's available to all of us. When you see something broken in the tools you depend on, file the issue. When it gets ignored, come back. When it gets closed without a fix, say so, politely and specifically and in public. Build the workaround, share the workaround, keep the receipts. Companies respond to markets, and in open source, the issue tracker is the market.
The happy ending is real, for what it's worth. AGENTS.md support in Claude Code is well-implemented, configurable, and ships with the source visible. The docs even deprecate the community's own workarounds. That's what winning looks like, eventually.
It just took 13 months, 405 comments, and a community that wouldn't take "closed" for an answer.
Top comments (0)