AGENTS.md is a plain Markdown file, checked into a repo's root, that tells a coding agent what a human contributor already knows and an AI one doesn't: the build command, the test command, the formatting rules, which directories are off limits. It's aimed at the agent, not at people reading the repo on GitHub. On September 18, 2026, Claude Code started reading that file too, as a fallback for any project that doesn't already have a CLAUDE.md. That's a small change with a real consequence for anyone running more than one coding agent against the same codebase: Claude Code, Cursor, and GitHub Copilot no longer each need their own separate instruction file to get the same context.
What AGENTS.md actually is
AGENTS.md is a spec, not a Claude Code invention. According to agents.md's own definition, it's "a simple, open format for guiding coding agents," built to give a project "a dedicated, predictable place to provide the context and instructions to help AI coding agents work on your project." The same page credits the format's origin to an effort spanning "OpenAI Codex, Amp, Jules from Google, Cursor, and Factory," and says it's now stewarded by the Agentic AI Foundation under the Linux Foundation. Adoption isn't small: agents.md's own count puts current use at over 60,000 open-source projects.
The site also states its own precedence rule for the simple case: the closest AGENTS.md file to the one being edited wins, and an explicit chat prompt from a person overrides whatever the file says. That's clean when AGENTS.md is the only instruction file in a repo. It gets messier the moment a repo also carries a CLAUDE.md, a .github/copilot-instructions.md, or a .cursor/rules folder, which is common in repos that more than one agent works on.
Claude Code's September 18 change
Claude Code's changelog for version 2.1.277, dated September 18, 2026, describes the change in one line:
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)
Three things follow from that sentence. First, it's a fallback, not a merge: Claude Code doesn't combine the two files or layer AGENTS.md content on top of an existing CLAUDE.md. If a CLAUDE.md is present, Claude Code reads it and ignores AGENTS.md entirely. Second, the read-or-don't toggle isn't a separate flag; it lives under "Project instructions" inside the existing /config command, next to whatever else that screen already manages. Third, the release ships with a real gap attached.
The Bedrock, Vertex, and Foundry gap
The same changelog entry ends with a parenthetical that's easy to skim past: AGENTS.md support isn't yet available "on Bedrock, Vertex or Foundry." A team running Claude Code through Amazon Bedrock, Google Vertex AI, or Microsoft Foundry, rather than Anthropic's own API, doesn't get this release yet. Nothing in the changelog explains why or promises a date, and there's no reason to guess one. Treat it as a gap in v2.1.277, not a permanent split between deployment paths.
One shared file or one per tool? The real decision for a multi-agent repo
A repo running Claude Code, Cursor and GitHub Copilot at once fits the mixed-tool policy JetBrains' adoption data is pushing teams toward. That kind of repo has an actual decision to make: write one AGENTS.md all three tools can read, or keep a separate instruction file per tool.
Put each tool's documented behavior next to the others and the gap is obvious. Claude Code reads AGENTS.md only when there's no CLAUDE.md in the project; if both exist, CLAUDE.md wins outright and AGENTS.md never gets consulted. Cursor's CLI reads AGENTS.md and CLAUDE.md at the project root and applies them "as rules alongside .cursor/rules," per Cursor's CLI documentation. Cursor's own rules documentation lays out a separate Team Rules, Project Rules, and User Rules chain, but never says where AGENTS.md fits into it. GitHub Copilot's coding agent reads AGENTS.md, including nested files in subdirectories, but according to GitHub's changelog announcing the support, it also still reads .github/copilot-instructions.md, .github/instructions/**.instructions.md, CLAUDE.md, and GEMINI.md, with no stated order between any of them.
The docs are each accurate on their own terms: three vendors each solved instruction-file discovery for their own product without reconciling it against the other two. GitLab's own governance writeup already flags this exact fragmentation, though its version of the problem is permissions, not instructions: a team running Duo on one repo, Claude Code on another, and Cursor on a third, each with a different permission model and no shared source of truth. Instruction files carry the same shape of problem one layer earlier: before an agent decides what it's allowed to touch, it first has to work out which file is telling it how to behave at all. That's a different axis from GitHub's separate deny-ask-allow permission model, which governs what an agent may do, not what instructions it reads before doing it.
For a repo already running more than one agent, one AGENTS.md at the root, with tool-specific files only where a tool needs something the others don't, beats maintaining three copies of the same conventions that drift the first time someone updates one and forgets the other two. Cursor's own coordinator agent, which now runs unprompted inside a Project, is one more reason to keep that file current: an agent that starts work without a person triggering it is reading whatever instructions happen to be in place, not whatever a developer meant to write down.
What to put in an AGENTS.md file
AGENTS.md isn't a second README. The convention agents.md itself promotes is content a human contributor wouldn't need but an agent does: the exact commands that build the project and run its test suite, formatting and linting rules, naming conventions, anything that would just clutter a README aimed at people. A file that's a copy of the README with different capitalization isn't doing the job.
That distinction matters more once a repo has picked the one-file approach from the section above. An AGENTS.md written for three different agents needs to be specific enough that none of them has to guess: the actual test command, not "run the tests," and the actual lint rule that fails CI, not "follow best practices."
The governance behind the standard
AGENTS.md isn't controlled by any one vendor. The Linux Foundation announced the formation of the Agentic AI Foundation on December 9, 2025, in the same press release stating that AGENTS.md was released by OpenAI in August 2025 as "a simple, universal standard that gives AI coding agents a consistent source of project-specific guidance." Anthropic, Block, and OpenAI are named as founding contributors to the foundation itself.
The same announcement folded in a second donation: MCP moved into the same foundation on the same day. Anthropic's Mike Krieger described donating MCP to the Linux Foundation as a way to keep it "open, neutral, and community-driven." Two of the protocols an AI coding agent depends on most, what it's allowed to call and what it's told to do, now sit under the same non-profit steward instead of inside a single vendor's roadmap.
FAQ
What is AGENTS.md?
An open, plain-text file that gives a coding agent project-specific instructions: build and test commands, conventions, anything that would clutter a README meant for people. It's used by more than 60,000 open-source projects and is stewarded by the Agentic AI Foundation under the Linux Foundation, not by any single vendor.
Does Claude Code read AGENTS.md if I already have a CLAUDE.md?
No. CLAUDE.md wins whenever it exists. Claude Code only falls back to AGENTS.md in a project that has no CLAUDE.md at all, and it doesn't merge the two files or layer AGENTS.md content on top.
Which AI coding tools support AGENTS.md?
Claude Code, since September 18, 2026, as a CLAUDE.md fallback. Cursor's CLI, which reads AGENTS.md and CLAUDE.md at the project root alongside its own .cursor/rules. GitHub Copilot's coding agent, including nested AGENTS.md files in subdirectories, on top of the several other instruction formats it already reads. agents.md's own list names other adopters, including OpenAI Codex, Jules, Amp, and Factory, though not every tool on that list documents its own precedence rule the way these three do.
Where do I turn AGENTS.md support on or off in Claude Code?
Under "Project instructions" inside Claude Code's /config command, per the same changelog entry that introduced the feature. There's no separate flag outside that screen.
Does AGENTS.md work with Claude Code on Bedrock, Vertex, or Foundry?
Not as of version 2.1.277. Anthropic's changelog states plainly that the feature isn't yet available on Bedrock, Vertex, or Foundry, without giving a reason or a timeline. Teams running Claude Code through those platforms don't get the fallback yet, whatever their CLAUDE.md setup looks like.