AGENTS.md emerges as standard for AI coding agent instructions

The AGENTS.md format has emerged as a standard for briefing AI coding agents, currently utilized by over 60,000 open-source projects and stewarded by the Agentic AI Foundation under the Linux Foundation. While the format is effective for providing stable operational facts—such as installation steps, test commands, and pull request formats—critics argue it serves more as a workaround than a complete solution for complex governance.

An analysis of 558 AGENTS.md files reveals that most files are written defensively, focusing on boundaries and build/test commands. Specifically, 85.7% of files list prohibitions, and 82.8% include build or test commands. However, there is a significant lack of information regarding 'gotchas' (pitfalls not derivable from code), which appear in only 13.6% of files, and 'agent_meta' (rules regarding the agent's tone or behavior), which appear in 25.8% of files.

While the format excels at providing context that would be given to a human teammate, it faces structural limitations when tasked with managing evolving architecture, security postures, and complex coding standards.