Agent Skills from the Perspective of Procedural Memory: A Survey

As large language model (LLM)-powered autonomous agents increasingly demonstrate strong capabilities in complex real-world tasks, recent research has focused on leveraging skills to endow these agents with domain-specific expertise and adaptive competence. In this survey, we present the first systematic overview of LLM-powered Agent Skills from the perspective of procedural memory, conceptualizing skills as a core form of procedural knowledge that enables agents to internalize, retain, and reuse task-solving procedures through interaction and experience. We comprehensively examine what Agent Skills are, why they are essential for agentic intelligence, and how they can be effectively acquired, represented, invoked, and refined within LLM-based systems. Furthermore, we review representative applications across diverse domains and discuss key challenges and future opportunities in developing and generalizing Agent Skills.