Open a repository that several developers touch, and you may find more than one instruction file. CLAUDE.md for Claude Code. AGENTS.md for Codex, Cursor, and others. They start as copies. Then they drift. Then an agent follows a build command that stopped working a month ago.

Many teams patched this with a symlink. It works. It’s also one more thing to maintain.

On September 18, Anthropic shipped Claude Code 2.1.277, and it changes the picture. Thariq Shihipar of the Claude Code team announced it this way: “We’re adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md.”

Whose spec is it, anyway? The headlines call it OpenAI’s, and that’s true of its origin. But OpenAI contributed AGENTS.md to the Agentic AI Foundation, part of the Linux Foundation, when it launched in December 2025. Anthropic contributed the Model Context Protocol to the same foundation. Both companies are founding platinum members. The AGENTS.md site says more than 60,000 open-source projects use the format and lists more than 30 compatible tools. So this looks less like one vendor bowing to another. It’s a shared standard reaching a tool many teams use every day.

Mitch Ashley, vice president and practice lead for CIO & technology buyers and software lifecycle engineering at The Futurum Group, says that’s the point. “For most enterprises, running several coding agents at once is now the permanent state, so convergence on one instruction format matters more than which vendor adopted it.”

Here’s what the behavior looks like in practice, based on Anthropic’s documentation. By default, Claude Code reads AGENTS.md only when there’s no CLAUDE.md in the working directory or any directory above it. If both files exist, Claude reads CLAUDE.md and ignores AGENTS.md. If a CLAUDE.md already imports AGENTS.md, Claude loads it through the import.

The “Project instructions” setting in /config has four values. The default is claude-md-or-agents-md. There’s also claude-md-and-agents-md, which loads both, CLAUDE.md first. Then claude-md, which ignores AGENTS.md entirely, and managed-only, which loads only the organization’s managed CLAUDE.md and auto memory.

Monorepos get sensible treatment. When Claude opens a file in a subdirectory with the Read tool, it picks up that folder’s AGENTS.md too, as long as the folder has no CLAUDE.md of its own. That matches the AGENTS.md convention of placing a file in each package, with the nearest one taking precedence.

Now the caveats. DevOps teams will want to read these before they delete any symlinks.

First, CLAUDE.local.md counts as a CLAUDE.md. A developer who adds one for personal preferences will stop Claude from reading the team’s AGENTS.md. The documented fix is the claude-md-and-agents-md setting.

Second, support is narrower than the file name suggests. Claude does not read AGENTS.local.md, AGENTS.override.md, or anything under a .agents/ directory. Teams that keep skills there will need another route.

Third, availability. The changelog says the feature isn’t on Bedrock, Vertex or Foundry yet. The documentation adds more conditions. Sessions that don’t fetch feature flags from Anthropic can’t read AGENTS.md, which includes sessions with telemetry disabled. The first session after an install or upgrade doesn’t read it either. In those cases, the docs say to import AGENTS.md from a CLAUDE.md instead. Teams that install Claude Code fresh on ephemeral CI runners should test this before they rely on it.

Fourth, visibility. An AGENTS.md loaded through the setting doesn’t appear in /memory or in the memory files list in /context. InstructionsLoaded hooks don’t fire for it either. They do fire for an AGENTS.md that a CLAUDE.md imports or symlinks to. If your governance tooling watches those hooks, that difference matters.

None of this is a reason to hold off. It’s a reason to check your setup once.

The larger point is about what these files are. An instruction file tells an autonomous agent which commands to run, which to avoid, and how your team builds software. That’s configuration, and Ashley sees a gap in how it’s handled. “These files are the first governance artifact a team has for agent behavior, and they sit outside the review path that already covers pipeline configuration.”

Three drifting copies widen that gap. One shared file gives reviewers one place to look. Ashley is specific about who should be looking. “Give them an owner in platform engineering.”

If you run more than one coding agent, a short checklist will do. Count the instruction files in your repositories. Pick AGENTS.md as the source of truth. Look for CLAUDE.local.md files that could hide it. Then confirm it loads on the platform you actually use, whether that’s the CLI, a CI runner, or a cloud provider.

Ashley adds one caution about what a shared file can’t do. “The harder interoperability problem is agent context and orchestration across vendors, and no shared file fixes that one.”

Anthropic didn’t invent this fix. It adopted one that a lot of other tools already use. For teams that maintain these files, that’s the useful part.