Anthropic released Claude Code 2.1.277 (September 18, 2026): if a project has no CLAUDE.md file, the agent now reads AGENTS.md, and operating modes are switched via the “Project instructions” section in /config. The feature was packaged as the first public “mod” for Claude Code, and with it, a single instruction file at the repository root is now understood by all five major agents — Claude Code, Codex, Cursor, Gemini CLI, and GitHub Copilot. A significant caveat: at launch, the fallback does not work on Amazon Bedrock, Google Vertex AI, or Microsoft Foundry.

image
image
image

What happened

On September 18, 2026, Anthropic released Claude Code 2.1.277. Instruction mode management is located in the “Project instructions” section of /config, where four options are now available: CLAUDE.md only, AGENTS.md as a fallback when CLAUDE.md is absent (the default), both files at once, and organization-managed instructions only. The feature was implemented as the first public “mod” (plugin) for Claude Code — the source code and documentation are published in the anthropics/claude-code repository under the mods/agents-md directory.

Context

AGENTS.md is an open instruction format for AI agents that OpenAI transferred to the Linux Foundation in December 2025 under the Agentic AI Foundation. The format is already supported by Codex, Cursor, Gemini CLI, and GitHub Copilot, and is used in more than 60,000 repositories. Before the 2.1.277 release, Claude Code users bridged the two worlds with workarounds: symlinks and one-line @-imports of AGENTS.md from CLAUDE.md — such workarounds broke in CI/CD. Technically, substituting AGENTS.md for CLAUDE.md is a change to prompt and context assembly, not a new model capability.

Why this matters for the industry

Until this release, Claude Code remained the last major agent without AGENTS.md support, so a single instruction file at the repository root now covers all five leading agents for the first time. For teams maintaining multi-agent repositories, this eliminates an entire class of desynchronization between vendor-specific instruction files and reduces support costs today. The mechanism itself is also engineering-significant: the “mod” is the first public groundwork for a Claude Code plugin layer, and the trust hierarchy is asymmetric: an organization can pin a mode via managed settings, but a repository’s inspectable .claude/settings.json cannot. Because the implementation is open source, the release can be verified by code rather than marketing claims.

Why this matters for users

Update Claude Code to 2.1.277 and remove the symlink or one-line @-import of AGENTS.md from CLAUDE.md — the file will be picked up automatically if the project has no CLAUDE.md. The “both files at once” mode lets you keep CLAUDE.md for Claude-specific rules while making AGENTS.md common to all agents. Note that the default mode is fallback: in projects without CLAUDE.md, the instruction set changes immediately after the update, so after upgrading it is worth running existing checks to ensure the agent reads what is expected. If you run Claude Code through Amazon Bedrock, Google Vertex AI, or Microsoft Foundry, behavior there does not change — continue keeping CLAUDE.md.

What is still unknown / limitations

Format portability is not the same as behavior portability: from open sources, it does not follow that different agents interpret the same AGENTS.md identically, and there are no benchmarks or claims of new model capabilities in the sources — from an eval standpoint, there is nothing to evaluate. There are also no GA dates for the “mod” contract or roadmap in the sources, nor any warnings or logs of fallback triggering; predictions about third-party mods and a separate .agents/ catalog remain interpretations, not facts.

Sources

Author

Look at AI, editorial team