large-language-model-agents

This paper introduces W2S, a framework that automatically constructs executable Skills for LLM agents from historical interaction traces using a Skill-IR intermediate representation, improving behavioral replay consistency by 10.5% over baselines.