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 become a common standard read by Codex CLI, GitHub Copilot,…

The rise of AGENTS.md as a shared configuration format was supposed to end a period of fragmentation in AI coding tools, where every agent expected its own instructions file. According to a developer's account documented on Zenn, the format has largely succeeded at that goal, yet a simple audit of a single work laptop turned up 111 separate AGENTS.md files. The finding matters because it suggests that standardizing the file format does not, by itself, address the deeper problem of configuration sprawl and maintenance overhead.

AGENTS.md is a plain-text file placed in a repository to give an AI coding agent context: project conventions, build commands, testing instructions, and other guidance the model should follow. Its appeal is convergence. The author notes that Codex CLI, GitHub Copilot, Cursor, Windsurf, Amp, and Devin all read the same file, which removes the need to duplicate instructions across tool-specific formats. A handful of exceptions remain, notably Claude Code, which continues to look for CLAUDE.md, and Gemini CLI, which uses GEMINI.md. The author argues that the list of holdouts keeps getting shorter, framing AGENTS.md as the emerging default rather than one option among many.

The central observation is that the format war was never the expensive part. To illustrate, the author counted the files on one ordinary machine, describing it as an unremarkable setup: a single developer, several years of accumulated projects, and a normal agent environment rather than anything unusual or contrived. The count reached 111 AGENTS.md files on disk. The account also distinguishes between the raw number of files and how many actually differ in content, implying that a significant share may be near-duplicates or copies that have drifted apart over time. That distinction is important, because it points to redundancy and inconsistency as the real cost rather than the file syntax itself.

This pattern will be familiar to anyone who has managed configuration at scale. When each repository, and sometimes each subdirectory, carries its own instructions file, the total set grows quietly with every new project. Over years, copies accumulate, get lightly edited, and slowly diverge, leaving a developer with many files that were meant to say the same thing but no longer do. Unifying on a single filename solves the question of which tools can read the file, but it does not answer how those files stay synchronized, who owns the canonical version, or how updates propagate across dozens of projects.

The situation echoes long-standing challenges in developer tooling. Configuration files such as .editorconfig, .eslintrc, .prettierrc, and .gitignore have faced similar dynamics, where teams often maintain shared templates, inherit settings hierarchically, or pull common rules from a central package to avoid drift. Some organizations manage this with dotfile repositories, monorepo-wide defaults, or generated configuration. AGENTS.md, as a relatively new artifact tied to the fast-moving AI agent ecosystem, does not yet have an equally mature set of conventions for inheritance or deduplication, which appears to be part of why the files multiply.

There is also a broader industry backdrop. Agent-oriented development tooling has expanded rapidly, with vendors competing to embed AI assistants directly into editors, terminals, and CI pipelines. Interoperability efforts like a shared instructions file are part of a wider push toward common interfaces, comparable in spirit to other attempts to standardize how models receive context and connect to tools. Convergence on a single file is a meaningful step, but the laptop audit is a reminder that standard adoption and operational simplicity are not the same thing.

For teams evaluating how to organize agent instructions, the takeaway is likely to be practical rather than alarming. A single shared format reduces friction between tools, and that benefit is real. The harder work is governance: deciding where the authoritative instructions live, whether project-level files should reference a shared source instead of copying it, and how to prune or reconcile the copies that already exist. Until those practices mature, the number of AGENTS.md files on a given machine may keep climbing, even as the format debate itself appears to be settling.