CUA-Skill: Develop Skills for Computer Using Agent
Tianyi Chen, Yinheng Li, Michael Solodko, Sen Wang, Nan Jiang, Tingyuan Cui, Junheng Hao, Jongwoo Ko, Sara Abdali, Suzhen Zheng, Leon Xu, Hao Fan, Pashmina Cameron, Justin Wagle, Kazuhito Koishida
TL;DR
CUA-Skill addresses the lack of reusable, human-aligned skill abstractions for desktop CUAs by introducing a structured skill library of parameterized execution graphs and composition graphs, plus a retrieval-augmented CUA-Skill Agent. The framework enables scalable skill expansion, memory-aware failure recovery, and robust long-horizon task completion, achieving state-of-the-art results on WindowsAgentArena (best-of-three $57.5\%$) and strong trajectory generation performance ($1.7\times$–$3.6\times$ gains). Across 17 Windows apps, 452 atoms skills yield $76.4\%$ average skill success with modest per-skill actions, illustrating reliable executability and generalization. The results demonstrate the practical viability of a skill-centric desktop agent substrate that improves robustness and efficiency across LLM backbones, paving the way for scalable, reusable, and interpretable CUAs.
Abstract
Computer-Using Agents (CUAs) aim to autonomously operate computer systems to complete real-world tasks. However, existing agentic systems remain difficult to scale and lag behind human performance. A key limitation is the absence of reusable and structured skill abstractions that capture how humans interact with graphical user interfaces and how to leverage these skills. We introduce CUA-Skill, a computer-using agentic skill base that encodes human computer-use knowledge as skills coupled with parameterized execution and composition graphs. CUA-Skill is a large-scale library of carefully engineered skills spanning common Windows applications, serving as a practical infrastructure and tool substrate for scalable, reliable agent development. Built upon this skill base, we construct CUA-Skill Agent, an end-to-end computer-using agent that supports dynamic skill retrieval, argument instantiation, and memory-aware failure recovery. Our results demonstrate that CUA-Skill substantially improves execution success rates and robustness on challenging end-to-end agent benchmarks, establishing a strong foundation for future computer-using agent development. On WindowsAgentArena, CUA-Skill Agent achieves state-of-the-art 57.5% (best of three) successful rate while being significantly more efficient than prior and concurrent approaches. The project page is available at https://microsoft.github.io/cua_skill/.