This STUDY Changes EVERYTHING We Know About AGENTS.md
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
An ETH Zurich study found that AGENTS.md files (used as AI coding agent instructions) often make agents perform worse, not better. Three main reasons: overgeneralization (broad rules applied where inappropriate), redundancy (agents can discover project structure themselves), and outdated content (instructions drift out of sync with the codebase). The recommendation shifts to keeping AGENTS.md concise and globally applicable, using a three-question audit: Does this apply to every task? Can the agent figure it out alone? Is this too specific for a global rule? Specific instructions should be moved to scoped rules or agent skills, though agent skill self-discovery is currently unreliable beyond three skills.