A new post from Augment examines how `AGENTS.md` files appear to affect code-generation quality inside its monorepo, and the results seem to be mixed in a fairly dramatic way. According to the company’s writeup, some guidance files noticeably improved agent output, while others made results worse than having no `AGENTS.md` at all.
The broader lesson, as the post suggests, is that the best files are narrow, structured, and easy for an agent to follow without wandering into unrelated material. Augment highlights patterns such as concise workflows, decision tables, and selective examples from production code, while warning that overly broad architecture notes or long lists of “don’ts” can push the model into unnecessary exploration. The full post goes deeper on how the team measured those effects and how it thinks existing docs can be reorganized for better results.
Source: Augment