I run my company on a fleet of AI agents. One answers the phones. One works new leads like a sales rep. Others build websites, write content, and keep the reporting straight. Every one of them leans on a plain text file that tells it how we do things here: our brand, our rules, what it’s allowed to touch, when to stop and ask a human.

That file is the whole game. The agent is only as good as the instructions behind it.

And for a long time, that file only spoke one dialect.

Mine were written for Claude Code, in a file called CLAUDE.md. If I wanted to move one of those agents to a different tool, say OpenAI’s Codex or Cursor, I couldn’t just carry the instructions over. Each tool wanted its own file, in its own spot, named its own way. Same instructions, rewritten three times, kept in sync by hand. Miss one, and an agent goes off-script in front of a customer.

I lived with it. Most people building on AI right now are living with some version of it and figuring it’s just the cost of doing business.

Last week it changed.

Claude Code started reading a file called AGENTS.md. Here’s why that one line matters: AGENTS.md isn’t Anthropic’s file. It’s a shared standard that Codex, Cursor, Copilot, Windsurf, and about twenty other tools already read. It came out of the teams behind Codex, Cursor, and Google’s Jules, who all got tired of inventing five names for the same thing. It’s now run by a foundation under the Linux Foundation, which means no single company owns it.

So when Claude Code, the most widely used coding agent of the bunch, started reading it too, that’s the moment a convention turned into a standard.

I’m not going to oversell it. It’s not full parity yet. CLAUDE.md still wins if both files are sitting there. Claude Code only reaches for AGENTS.md when there’s no CLAUDE.md, and the support is even switched off on a couple of the big enterprise setups for now. If you read the headline and thought “one file, everywhere, done,” it’s not that clean.

But I don’t care much about today’s exact behavior. I care about the direction. The biggest name in the space just agreed to read everyone else’s file. That doesn’t un-happen.

So here’s what my team is doing about it.

We’re moving every one of our agents onto AGENTS.md as the real source of truth. The old CLAUDE.md files become a single line that points at it. One set of instructions per agent, written once, read by whatever tool we point at the job. If a better coding agent shows up in six months, and one will, we move to it in an afternoon instead of a rewrite.

That’s the part I want you to take, whether you write a line of code or not.

The AI tool you pick today is not going to be the best tool a year from now. This stuff moves too fast. If your whole setup is wired into one vendor’s format, one vendor’s file, one vendor’s way of doing things, you’ve quietly signed up to rebuild it every time something better comes along. Or worse, you stay on the weaker tool because leaving costs too much.

The move is to build on what you can carry. When a real standard shows up, get on it early, before you’ve got fifty files locked to one vendor and no clean way out. Portability isn’t a technical nicety. It’s the difference between the market working for you and you scrambling to keep up with it.

I’ve been telling clients this for a while, mostly getting nods and not much action. It’s easier to point at now. The standard is here. The biggest player just joined it.

So go look at what you’ve built on AI this year and ask one question: if the tool underneath it disappeared tomorrow, how hard would it be to move? If the answer makes you wince, that’s your next project.