You don't write AGENTS.md. You train it with gradient descent. - View it on GitHub