AGENTS.md in the 📝Claude Code and Codex context is a markdown instruction file — analogous to 📝CLAUDE.md as Infrastructure — that cross-tool agents read at session start, enabling shared project conventions across multiple AI coding tools.
AGENTS.md emerged as a cross-vendor convention for projects using multiple AI coding agents in the same repository. OpenAI's Codex CLI, Anthropic's Claude Code, Google's Gemini CLI, and several open-source tools all read an AGENTS.md file at the project root when they boot — giving a single place to encode project conventions, build commands, test patterns, and behavioral guidelines that apply regardless of which agent is running.
The file is structurally identical to CLAUDE.md or GEMINI.md — plain markdown with whatever sections the author finds useful — but the filename is the coordination mechanism. A project can ship CLAUDE.md, AGENTS.md, and GEMINI.md as separate files when each tool needs different instructions, or rely solely on AGENTS.md when the instructions are tool-agnostic and the goal is to write once for all agents.
In Claude Code, AGENTS.md is read alongside CLAUDE.md at session start, with CLAUDE.md taking precedence on conflicts. This lets practitioners use AGENTS.md for the shared core and CLAUDE.md for Claude Code-specific overrides. Adoption has been driven less by formal specification and more by emergent convention as multi-agent workflows proliferate — note that a separate AGENTS.md convention exists in 📝OpenClaw as a sub-agent registry, an unrelated naming collision worth being aware of.