Just two months ago in my last talk I mentioned that there's an AGENT.md approach from some enthusiastic folks who formalized it as a manifest/proposal to agent developers - to unify the rules. Otherwise every agent has its own format, everything is in different places, and so on. But the essence is the same - give the agent a base prompt.
It was nothing more than a proposal back then. You still had to put files that each tool expects, but the complexity of maintaining was solved with symlinks from them to this one AGENT.md.
And literally yesterday I discovered that AGENT.md redirects to AGENTS.md (with an S at the end, naming is hard). And this is no longer a proposal from some random dudes, but a standard from OpenAI. And, even more surprisingly, it's already been supported by almost everyone in this time: Codex, Roo, Cursor, Aider, opencode, zed, and many others - even Copilot, which doesn't rush to add new features.
Looking contemptuously at Anthropic. Claude Code is almost the only one that doesn't have this support yet.
The direction is definitely pleasant and good. But it feels like they didn't quite finish it. You want some sufficiency from a "standard", not minimalism. For example, Claude Code has: the ability to decompose into different files and inline them into each other via @, the ability to divide rules into general and personal, the ability to set rules at the system level. Or Roo also did it well - there it's just initially a directory with all this stuff, you can immediately split the base prompt into several files. But this standard has nothing but AGENTS.md in the root.
It's very annoying, by the way, that every first tool and standard pushes you to keep all this in the project root. Most often you can't even keep it anywhere else - it won't work. And everyone has different naming principles. From .gitignore and .editorconfig, to CLAUDE.md, Dangerfile.df.kts and .gitlab-ci.yml. At least you can change the latter.