AGENTS.md: A Cross-Vendor Standard for Instructing AI Coding Agents

AGENTS.md is an open, Markdown-based format for giving AI coding agents project-specific instructions; originated by OpenAI, it is now stewarded by the Linux Foundation's Agentic AI Foundation and, as of September 2026, is read natively by Claude Code as well as Codex, Gemini CLI, Cursor, and dozens of other tools.

Created 2026-09-19 Last reviewed 2026-09-19

What it is

AGENTS.md is a plain-Markdown file that developers place at the root of a code repository to tell AI coding agents how to work in that project — what commands to run to build and test the code, which coding conventions to follow, which directories to avoid, and other operational context that would clutter a human-facing README. It has no required fields or proprietary syntax; it is simply Markdown that a coding agent reads before acting.

The format grew out of a convention originally associated with OpenAI’s Codex agent and was formalized through a collaborative effort involving OpenAI, Google’s Jules and Gemini CLI teams, Cursor, Sourcegraph’s Amp, and others. Since its release, adoption has been extensive: the format’s own documentation and independent press coverage put usage above 60,000 open-source projects and repositories, with support built into tools including GitHub Copilot’s coding agent, Devin, JetBrains Junie, Aider, goose, Zed, Warp, and VS Code.

In practical terms, AGENTS.md solves a coordination problem: a team using several different AI coding tools no longer needs to maintain a separate instruction file for each one. One file, read by all.

Why it matters for AI governance and narratives

AGENTS.md is a small but telling data point in a larger pattern this observatory tracks: firms that compete aggressively at the model layer — on benchmark performance, pricing, and safety claims — often converge quickly on shared technical infrastructure beneath that competition. The same week vendors’ AI models were reportedly breaching one another’s guardrails in adversarial testing, their coding-agent products aligned on a common configuration format. That is not a contradiction; it is a familiar pattern in software markets, where interoperability standards lower switching costs and expand the addressable market for everyone, even as branding and marketing continue to emphasize differentiation.

The governance dimension is the venue chosen for stewardship. OpenAI did not keep AGENTS.md as a wholly owned de facto standard; it contributed the specification to the Agentic AI Foundation (AAIF), a vendor-neutral body under the Linux Foundation, alongside Anthropic’s Model Context Protocol (MCP) and Block’s goose framework. Placing competing companies’ contributions under one neutral foundation is a governance move as much as a technical one — it signals an attempt to establish shared, non-proprietary rails for agentic AI infrastructure before any single vendor’s format could be positioned as a competitive moat or, conversely, before fragmentation made cross-tool AI development unworkable. Anthropic’s decision to build native AGENTS.md support into Claude Code is best read in that light: a company adopting a rival’s contributed standard because the coordination benefits outweigh the cost of not controlling the format Claude Code reads by default.

Key facts and dates

OpenAI released AGENTS.md as an open specification in August 2025. On December 9, 2025, the Linux Foundation announced the formation of the Agentic AI Foundation, with three founding project contributions: Anthropic’s Model Context Protocol, Block’s goose, and OpenAI’s AGENTS.md. Founding platinum members of the foundation included Amazon Web Services, Bloomberg, Cloudflare, Google, and Microsoft.

On September 18, 2026, Anthropic shipped Claude Code version 2.1.277 with native, automatic support for AGENTS.md: when a repository contains no CLAUDE.md file (Claude Code’s own, longer-standing instruction format, which predates AGENTS.md), Claude Code now falls back to reading AGENTS.md instead. The behavior is conservative — a repository that already has a CLAUDE.md is unaffected, and teams can still layer CLAUDE.md on top of a shared AGENTS.md for tool-specific tuning via an import mechanism. As of that date, the fallback applied to the standalone Claude Code CLI and desktop application but had not yet extended to cloud-hosted deployments such as AWS Bedrock, Google Vertex AI, or Anthropic Foundry.

Where to learn more

Sources

Primary source: the format's own canonical documentation, describing purpose, syntax, governance under the Agentic AI Foundation, and adopting tools
Primary source: the Linux Foundation's own press release naming founding contributors (Anthropic, Block, OpenAI) and platinum members
Reputable independent technology press covering Anthropic's September 18, 2026 decision to add AGENTS.md support to Claude Code
Secondary corroborating coverage confirming the version number (2.1.277), release date, and fallback-only behavior
Referenced in: Editorial No. 329