New survey defines 'agent skills' as key to scalable AI agents

A new arXiv survey from Zhou, Shu, Su, Du, Fang, and Lin tackles a critical bottleneck in LLM-based agents: the inefficiency of from-scratch reasoning and low-level tool calls for every task. The authors propose the concept of *agent skills*—reusable procedural artifacts that coordinate tools, memory, and runtime context under task-specific constraints. This shifts the paradigm from agents doing everything themselves to a division of labor: agents handle high-level reasoning and planning, while skills provide reliable, reusable, and composable execution. The paper cites real-world systems like OpenClaw and Claude Code as examples of this shift toward action-oriented task execution.

The survey systematically organizes the literature around four lifecycle stages of agent skills: representation (how skills are encoded), acquisition (how they are learned or programmed), retrieval (how relevant skills are selected at runtime), and evolution (how skills are updated and improved). Across these stages, the authors review methods, ecosystem resources (e.g., tool libraries, skill repositories), and application settings from software automation to robotics. They also identify open challenges: quality control to prevent faulty skills, interoperability across different agent frameworks, safe updating without breaking existing workflows, and long-term capability management as the skill library grows. The paper aggregates over 100 research papers, datasets, and open-source projects, providing a valuable roadmap for practitioners building next-generation agent systems.