GUI agents that operate across long-horizon tasks face a persistent problem: pop-ups, delayed page loads, and relocated widgets routinely break plans that were fixed before execution began. Existing agent-skill frameworks try to address this by packaging reusable procedural knowledge, but according to a new paper posted on Hugging Face, those skills are typically designed without targeting GUI execution dynamics and are treated as static artifacts produced before deployment.
The paper argues that what GUI agents need is not better static skills, but skills that can be revised from execution feedback at deployment time, without additional training. To that end, the authors propose EvoSkill-GUI, a training-free framework in which each skill is a structured multi-file package containing retrieval metadata, executable plans, backup localization, failure-recovery rules, accessibility utilities, and failure cases.
EvoSkill-GUI operates through a reflect-revise-reuse loop. The executor performs instant in-rollout revisions, an isolated critic diagnoses failed trajectories under strict information isolation, and the executor then edits specific skill files through a restricted tool interface.
The framework was evaluated across three mainstream GUI benchmarks spanning mobile and desktop platforms: MobileWorld, AndroidWorld, and OSWorld. According to the paper, EvoSkill-GUI consistently improved multiple base models without any training, with maximum gains of +16.2%, +6.0%, and +10.5% respectively. The authors also report that evolved skill libraries continue to benefit related tasks rather than being rebuilt from scratch.
Code is available at https://github.com/ZJU-REAL/EvoSkill-GUI.