Regression tax for agent skills

The arXiv paper 'The Regression Tax' studies nearly 6,000 office-automation runs and finds that adding procedural skills can improve average agent performance while still making some tasks worse. Why it matters: Skill libraries and saved workflows need rollback tests, not only success stories. A better average can hide a broken edge case in the review queue.

Try this: When adding an agent skill, keep a small regression set: one task it should improve, one stale-context case, and one case where the skill should stay out of the way.