Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents

TL;DR - This paper analyzes how reusable skills can cause LLM-agent failures and higher execution costs. Across two benchmarks, it identifies 307 skill-induced regressions and introduces SkillTriage for evidence-based failure attribution.

• The study finds 125 functional failures and 182 efficiency regressions.
• Seemingly relevant skills can cause agents to implement requirements incorrectly or omit them.
• Efficiency regressions are not attributable to prompt length alone.
• Excessive verification and heavy implementation pipelines account for 67 and 30 excessive-procedure cases, respectively.