Agent Skills Paper Counts 307 Failures

The paper punctures the assumption that more guidance is automatically beneficial: skills can quietly turn optional advice into costly, failure-inducing procedure.

• –Relevant skills—not just irrelevant ones—often cause agents to omit requirements or implement the wrong behavior
• –Excessive verification and heavyweight implementation pipelines account for 67 and 30 efficiency regressions, respectively
• –Prompt length alone does not explain the observed slowdowns and token costs
• –Skill libraries need outcome-based evaluation, selective loading, and lifecycle governance
• –SkillTriage offers a practical path toward debugging skills like production dependencies