Since December 2025, AGENTS.md has been stewarded by the Agentic AI Foundation under the Linux Foundation — and adoption has followed. The file, a plain markdown document that lives in a repository and tells an AI agent how to run tests, which style to follow, and where not to touch, is now read natively by OpenAI Codex, Cursor, GitHub Copilot’s coding agent, Windsurf, Amp, Aider, Gemini CLI, Zed, Jules and Devin. Over 60,000 open-source projects now carry one.
Why a config file matters more than it sounds
For most of 2024 and 2025, every AI coding tool wanted its own instructions file — a .cursorrules here, a tool-specific config there — meaning teams running more than one agent had to maintain duplicate, drifting sets of project rules. AGENTS.md fixes that by giving every tool a single shared source of truth: one file, read the same way regardless of which agent opens the repository.
That matters because of where the market has actually landed. As we covered on 1 July, engineering teams are increasingly running multiple AI coding tools on the same codebase rather than standardising on one vendor. A shared config format is the unglamorous plumbing that makes that workable at scale — without it, multi-tool workflows mean re-explaining your codebase’s conventions to every agent separately.
The holdouts are the tell
Not everyone signed on. Anthropic’s Claude Code still wants CLAUDE.md, and Google’s Gemini and Antigravity still want GEMINI.md. Both are the two most capable, most differentiated agents in the market — and both evidently see enough value in a proprietary config format to sit outside the open standard, at least for now. That’s worth watching: if the gap between “does what the open standard’s rules say” and “does what the proprietary format enables” stays meaningful, expect the holdouts to hold longer.
So what
If your development partner is running more than one AI coding tool against your codebase — which is increasingly the professional norm — ask whether their project conventions live in one format the whole stack can read, or are duplicated and drifting across tools. It’s a small operational detail that tells you a lot about how disciplined the engineering process actually is. More on how we run AI-assisted development day to day on the AI-assisted development page.