Definition
AGENTS.md is an emerging open convention: a plain-text, Markdown file that gives AI agents explicit instructions for how to understand, navigate, and act on a project, website, or business. It started in software engineering as a way to tell coding agents how a repository is structured, how to build and test it, and what conventions to follow, and the pattern is now extending to the web as a signal for autonomous agents that browse and transact.
Where llms.txt provides a curated map of a site's most important content for retrieval and citation, AGENTS.md is more action-oriented—it describes how an agent should interact: which workflows exist, what rules to respect, how to complete tasks, and where to find authoritative data. The two files are complementary supplemental signals rather than replacements for structured data or standard pages.
For coding agents, an AGENTS.md typically covers setup commands, test instructions, code style, and project structure so the agent can work productively without trial and error. For agent experience optimization, a site-level AGENTS.md can clarify business capabilities, supported actions, and constraints, helping agents recommend and transact correctly.
Like robots.txt and llms.txt, AGENTS.md is advisory: agents are not required to read or obey it. Its value lies in adoption by major agent platforms and in reducing ambiguity, which makes agents more likely to act on a business accurately rather than skip it.
Examples of AGENTS.md
- A software team adds an AGENTS.md with build, test, and lint commands plus code-style rules so coding agents can contribute correctly without guessing.
- A SaaS company publishes a site-level AGENTS.md describing supported actions—sign up, book a demo, check pricing—so AI agents route users to the right workflows.
- An ecommerce store pairs AGENTS.md with llms.txt: one maps key content for citation, the other explains how agents can browse the catalog and initiate checkout.
- A GEO team tests AGENTS.md adoption by checking whether agent platforms read the file and whether their instructions reduce mistakes in agent-driven tasks.
