Why Static Skill Inconsistencies Do Not Equate to Runtime Risks: A Behavioral Analysis of LLM Agent Execution

A systematic empirical study of 575 community and benchmark skills identified 1,182 manually validated contract inconsistencies, and a controlled evaluation pipeline with 49 executable agent tasks and three representative LLMs found that static defect volume cannot reliably reflect real-world agent performance degradation.