SkillX: Automatically Constructing Skill Knowledge Bases for Agents

Learning from experience is crucial for creating more capable LLM-based agents. However, the prevailing self-evolving paradigm is fundamentally inefficient: it forces each agent to learn in isolation, redundantly mining experiences from its own limited capabilities and scarce training data, resulting in expertise that fails to generalize. To break this cycle, we introduce SkillX, a framework that autonomously pre-builds a plug-and-play skill library. SkillX operates through a fully automated pipeline built on three synergistic innovations: **i) Multi-Level Skills Design**, which distills raw trajectories into three-tiered hierarchy of strategic plans, functional skills, and atomic skills; **ii) Iterative Skills Refinement**, which automatically revises skills based on execution feedback to continuously improve library quality; and **iii) Exploratory Skill Expansion**, which proactively generates and validates novel skills to expand coverage beyond seed training data. Using this framework, we construct a reliable plug-and-play skill library using a state-of-the-art agent, GLM-4.6. We conduct extensive experiments on challenging long-horizon, user-interactive benchmarks, including AppWorld, BFCL-v3, and $\tau^2$-Bench, demonstrating the effectiveness of SkillX. We also provide strategic insights into how experience transfer impacts model performance.