Claude Code 2.1.277 reads AGENTS.md as a fallback

Claude Code now reads AGENTS.md, an instruction file shared by several rival coding agents, when a project has no CLAUDE.md of its own. The change shipped on September 18, 2026 in version 2.1.277. One file can now brief more than one tool, which removes a small but real maintenance chore from any repository that uses two.

Anthropic's changelog states the rule in one line: "in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead; change it under 'Project instructions'". The matching GitHub release went out at 18:06 UTC the same day.

What AGENTS.md is

AGENTS.md is a plain Markdown file kept in a repository. It tells a coding agent how the project works, in the same way a contributing guide tells a new colleague.

The AGENTS.md site calls the format "a README for agents: a dedicated, predictable place to provide the context and instructions to help AI coding agents work on your project". It describes the contents as "build steps, tests, and conventions that might clutter a README or aren't relevant to human contributors".

The format is not Anthropic's own. It came out of work involving OpenAI Codex, Amp, Jules from Google, Cursor and Factory. It is now stewarded by the Agentic AI Foundation under the Linux Foundation. The project says more than 60,000 open-source repositories on GitHub carry the file.

How the fallback behaves

The lookup order is the whole feature, and it is short.

Item	Detail
Version	2.1.277
Released	September 18, 2026, 18:06 UTC
Rule	AGENTS.md is read only when the project has no CLAUDE.md
Toggle	"Project instructions" in `/config`
Not yet available on	Bedrock, Vertex, Foundry

Nothing changes for a repository that already has a CLAUDE.md. That file still wins, and the new behavior never merges the two. A project gets one or the other.

The fallback can also be turned off. Anthropic's changelog points to a switch under "Project instructions" in the /config menu.

One limit deserves attention before anyone plans around it. The changelog says AGENTS.md support has not reached Bedrock, Vertex or Foundry. Teams whose developers run Claude Code through a cloud provider will see different behavior from teammates running it directly.

The rest of the release

Version 2.1.277 was not only this feature. Anthropic's notes list more than 50 bug fixes across terminal rendering, file operations and MCP servers, the connectors that let an agent reach outside tools.

Two fixes stand out for anyone who has hit them. Permission prompts for dangerous removal commands now name the flagged command and suggest a ${VAR:?} guard, which makes an unset variable fail instead of deleting the wrong path. Servers set up as HTTP that still used the older HTTP and server-sent-events transport no longer fail on 422 and other 4xx responses.

A second release followed on September 19. Version 2.1.278 changed auto mode to default to a server-side classifier for Claude API and Enterprise users.

What this means for developers

Check which file your repository actually has before assuming anything changed. If a CLAUDE.md is present, this release does nothing for you, and an AGENTS.md sitting beside it stays unread.

For a project that supports contributors on several agents, the practical move is to keep one AGENTS.md and delete the CLAUDE.md that only duplicates it. That is now a supported setup rather than a workaround. Keep a CLAUDE.md only where you genuinely need instructions that differ per tool.

Watch for the split across your team. Because the cloud deployments do not have this yet, a contributor on Bedrock can get different behavior from one running the command-line tool, with no error to explain the difference. If your agent instructions matter to your review process, say in the repository which file is authoritative.

There is a wider point about how quickly agent behavior changes underneath you. Claude Code previously started writing a session link into commit messages unless you opted out, a default that reached repositories before many teams noticed. Configuration that lives in a checked-in file, reviewed like code, is the defense against both surprises.

Pin the version if your workflow depends on the old behavior. Releases here land daily, and a default that flips between two of them is easier to absorb when your team is all on the same build.

This article was first published on Tech AI Wire.