What is AGENTS.md?

AGENTS.md works like a README written for agents instead of human contributors. It sits in a repository and collects the information an agent needs to work there: how to set up and build the project, how to run tests, code style rules, and pull request or commit expectations. The project's site describes the format as having no required fields, so teams write whatever sections fit their codebase and treat the file as living documentation.

In a large monorepo each subproject can have its own AGENTS.md. Agents use whichever AGENTS.md is closest to the file being edited, and an instruction typed directly into the chat takes priority over every file. If the file lists test commands, an agent is expected to try running them and fix failures before finishing.

The site lists a long set of supporting tools, including OpenAI Codex, Google Jules, Cursor, GitHub Copilot's coding agent, Gemini CLI and VS Code. The site says stewardship of the format sits with the Agentic AI Foundation, which is part of the Linux Foundation. Teams that already use another instruction file name can rename the file and keep a symlink from the old name.