A Survey of Agent Skills: Toward Procedural Infrastructure for LLM Agents
Large language model (LLM) agents are increasingly expected to solve long-horizon tasks through repeated interaction with external environments, tools, users, and other agents. In this setting, agent skills have emerged as a central mechanism for transforming fragmented experience into reusable procedural knowledge. Unlike raw memory which preserves past traces, or abstract rules which often lack executable detail, skills compress recurring patterns of successful behavior into operational artifacts that can guide future action. However, existing studies on agent skills remain scattered across diverse areas, making it difficult to form a unified understanding of what skills are, how they are represented, and how they should be governed. To bridge this gap, this paper presents a comprehensive survey of agent skills. We propose a six-layer taxonomy covering ontology, representation and packaging, lifecycle, runtime integration, governance, and applications. We first define agent skills as reusable procedural abstractions that connect memory, human expertise, and execution. We then review major skill representations, including natural-language guidance, executable code snippets, decision graphs, filesystem-based packages, and structured skill records. Next, we analyze the skill lifecycle, from acquisition, storage, retrieval, usage, and maintenance to selective internalization into model behavior. We further examine how skills integrate with terminal interfaces, tool interfaces, multi-agent systems, and agent harnesses. In addition, we discuss the emerging skill ecosystem, its security and governance risks, and mechanisms for trusted deployment. Finally, we survey applications in robotics, games, web agents, GUI/mobile/OS agents, and software engineering, and identify open challenges in evaluation, robustness, standardization, safety, and infrastructure. Related resources, as well as the latest developments in this field, are accessible on https://github.com/DataArcTech/Awesome-Agent-Skill-Papers.
