AGENTS.md で統一したはずなのに、ノートPC 1台に AGENTS.md が 111 個あったWhile AGENTS.md has emerged as a shared configuration file across major AI…

AGENTS.md, a plain-text configuration file that tells AI coding assistants how to work inside a project, has quickly become one of the few points of agreement in an otherwise fragmented tooling landscape. Yet a recent blog post illustrates the flip side of that convergence: one developer discovered 111 separate AGENTS.md files scattered across a single laptop, a finding that underscores how a supposedly unified standard can still create real management overhead.

The premise behind AGENTS.md is straightforward. Instead of each AI coding tool inventing its own instruction file, the format offers a shared convention that multiple agents can read. According to the source, this has effectively ended the format war among several major tools: Codex CLI, GitHub Copilot, Cursor, Windsurf, Amp, and Devin all read the same file. That means a project can, in principle, document its build steps, coding conventions, testing commands, and other agent-facing guidance once, and have it apply consistently across different assistants.

Convergence is not total, however. Anthropic's Claude Code continues to use CLAUDE.md, and Gemini CLI still relies on GEMINI.md. In other words, teams that mix tools from different vendors may end up maintaining parallel files that contain overlapping or duplicated instructions. This partial adoption is central to why the "one file to rule them all" promise remains aspirational rather than fully realized.

An AGENTS.md file is typically written in Markdown and placed at the root of a repository, where it acts as a README aimed at machines rather than humans. It commonly captures things a coding agent needs to operate productively: how to install dependencies, how to run the test suite, which directories to avoid, and what style or architectural rules to follow. Because the convention supports nested files, larger monorepos can also place directory-specific AGENTS.md files closer to the relevant code, which is one reason a single project can contain more than one.

The count of 111 on a single machine likely reflects the sum of many factors rather than one project gone wrong. Developers routinely clone dozens of repositories, and dependency folders, sample projects, and vendored code can each carry their own AGENTS.md. As the convention spreads, the files multiply wherever code lives, and there is no central registry keeping track of them. The result is that a standard designed to reduce fragmentation can, paradoxically, increase the number of individual files a developer must be aware of.

This is a familiar pattern in configuration management. Formats such as .editorconfig, .gitignore, and the various ESLint and Prettier configuration files solved compatibility problems by standardizing syntax, but they also tend to accumulate across a filesystem over time. AGENTS.md appears to be following a similar trajectory, with the added wrinkle that its contents are instructions for autonomous or semi-autonomous agents, where inconsistent or outdated guidance can lead to unexpected behavior rather than a simple linting warning.

For teams adopting AI coding tools, the practical takeaway is that agreeing on a file format is only the first step. Keeping the actual content of those files accurate, deduplicated, and synchronized, especially across tools that still expect CLAUDE.md or GEMINI.md, remains a manual discipline. Some developers respond by symlinking or generating the vendor-specific variants from a single source, though the post does not present any one approach as definitive.

The broader context is an industry racing to make AI agents useful inside real codebases. As tools like Cursor, Copilot, and Codex CLI compete on capability, shared conventions such as AGENTS.md lower the switching cost between them and make projects more portable, which helps explain the format's rapid uptake. The 111-file discovery is a useful reminder that standardization at the format layer does not automatically deliver simplicity at the filesystem layer, and that the operational cost of these files can grow quietly as adoption scales. For now, AGENTS.md represents genuine progress toward interoperability, but the anecdote suggests that governance, tooling, and cleanup practices will need to mature alongside it before the promise of a single, unified configuration is fully met.