Coding agents are great at following patterns and terrible at guessing yours. The difference between an agent that ships and one that flails is almost never the model — it's whether the codebase tells it where the levers are.
Predictable paths beat clever prompts
If every feature lives in the same shape — table here, service there, hook here, page there — an agent can mirror the last one and get it right. No archaeology, no guesswork.
That's what an AGENTS.md at the repo root buys you: a control panel that says "to change the brand, edit this one file; to add a feature, mirror that one folder."
What it looks like in practice
- You prompt: "add a teams table with seat-based billing."
- The agent reads the manual, finds the reference feature, mirrors it.
- Types check, the app boots, the nav updates — first try.
The model didn't get smarter. The codebase got more legible. Build for the reader — even when the reader is a machine.