A study on arXiv measured what happens when a repository has an AGENTS.md file, the short plain-text document that tells a coding agent how the project is laid out, how to build it, and how to test it. Across ten repositories and 124 real pull requests (PRs), the file cut median agent runtime by 28.64% and output tokens by 16.58%, with task completion comparable to running without it.
The result makes sense to me. Without the file, the agent spends its first stretch of every task rediscovering the project: finding the test command, guessing the conventions, reading files to work out which ones matter. With the file, that work is done once, by a person who knows, and reused on every run.
What I find interesting is that this is the same document a good team already owes its new hires. How is the code organized, how do I run it, what are the rules, what should I never do. The agent is just a reader who cannot ask a colleague, so the gaps show up as wasted time instead of a question in chat.
So my view is that a repository-level instructions file is the cheapest technical standard a team can adopt. It takes an afternoon, it is reviewed like any other file, and it pays twice: once for every agent run and once for every person who joins. I would make it part of the definition of a repository, alongside the README and the lint config.
Does your repository have a file like this, and when was the last time someone read it?
Photo source: https://photos.robertstowe.com/venice