OSExpert: Computer-Use Agents Learning Professional Skills via Exploration

Revolutionizing Computer-Use Agents with Autonomous Skill Acquisition

OSExpert introduces a novel environment-learning paradigm for computer-use agents, moving beyond traditional scaled human demonstrations. By autonomously exploring digital environments, OSExpert agents acquire verifiable skill sets, achieve substantial performance gains (approx. 20% success rate increase), and significantly close the efficiency gap to human experts (by ~80%). The framework emphasizes bottom-up skill discovery, fast planning with skill boundary checks, and robust fine-grained action control through curated primitives, enabling agents to handle complex, long-horizon tasks and generalize to unseen UIs more effectively.