WeAreP2P Claude Code Now Reads AGENTS.md: The End of the Two-File Instruction Tax
Claude Code v2.1.277, released September 18, falls back to AGENTS.md when a project has no CLAUDE.md — ending the symlink era for developers who run multiple coding agents in one repo.
Developers who split their time between Claude Code and OpenAI’s Codex have been paying a small, persistent tax: two instruction files to maintain. CLAUDE.md for Anthropic’s agent, AGENTS.md for everything else. As of version 2.1.277, released September 18, that tax is gone — Claude Code now reads AGENTS.md whenever a project has no CLAUDE.md.
These instruction files are read by coding agents on essentially every request. They carry the standing context that shapes an agent’s behavior: build commands, testing requirements, coding conventions, preferred tool usage. Maintaining two of them meant either duplicating guidance across files or wiring up symlinks to keep CLAUDE.md and AGENTS.md in sync — a workaround that was widespread enough to be worth naming.
How the fallback works
The rule is deliberately simple. If a folder has no CLAUDE.md, Claude Code checks for AGENTS.md and uses it. CLAUDE.md keeps priority when both exist, so teams with Anthropic-specific instructions won’t see them silently replaced by the generic file. The behavior can be adjusted under “Project instructions” in /config, per the release notes. One limitation worth knowing: the fallback isn’t available yet on Amazon Bedrock, Google Vertex AI, or Microsoft Foundry, so teams running Claude Code through those providers are still on the old behavior.
The change was announced by Claude Code engineer Thariq Shihipar, and the response from developers who run several agents against the same repository was immediate relief. For multi-agent teams, the practical upshot is that one shared instruction file now reaches Claude Code without a Claude-specific duplicate. If you’ve been symlinking CLAUDE.md to AGENTS.md, you can delete the link.
Why Anthropic conceded the standard
AGENTS.md is OpenAI’s format, contributed to the Agentic AI Foundation under the Linux Foundation. More than 60,000 open-source projects had adopted it as of December 2025, according to The Register — a base large enough that holding out was costing Anthropic’s own users more than it was gaining the company. Claude Code’s popularity gave Anthropic the latitude to insist on its own convention, and for a while it did. Multi-agent repositories have now become common enough that the separatism reads as friction rather than differentiation.
As The Register put it, this detente is “an event of modest significance” — which is about right. Nobody’s workflow is transformed. But the direction of travel is clear: instruction files are converging on a shared standard the way tool protocols converged on MCP. The agent that reads your repo’s conventions without special-casing is the agent that fits into a team already running three other agents, and Anthropic just made sure Claude Code is one of them.