Computer-use agents can perform increasingly complex tasks in graphical interfaces, yet their interaction experience is typically fleeting: procedural knowledge acquired in one rollout is not systematically retained, refined, or reused in later tasks. Existing skill libraries supply external procedural knowledge, but their incremental benefit over the same agent without skills, and their longitudinal dynamics under repeated interaction, remain under‑characterized. We introduce an online skill‑evolution framework that turns interaction trajectories and evaluator feedback into a persistent, versioned library of reusable procedures. Each iteration runs against a frozen snapshot of the library, and evidence‑guided skill updates become available in subsequent iterations without altering model parameters. We evaluate the full evolving‑library system against a configuration‑matched empty‑library control across four OSWorld application domains, using the same fixed action‑generation and GUI‑grounding stack, task sets, and iteration horizons. After a five‑iteration empty‑library warm‑up, the full system achieves higher post‑warm‑up mean evaluator scores in all four domains, with differences ranging from 5.7 to 18.6 percentage points and domain‑dependent temporal stability. In the GIMP domain, provenance‑aware analysis reveals retrieval across task‑of‑origin boundaries and revision churn, where repeated accepted edits fail to recover the originating task. These findings characterize evolving skill libraries as auditable, shared procedural memory that can improve a fixed computer‑use stack, while showing that their benefits are conditional and repeated revision does not guarantee recovery. Code is released at https://github.com/LongtaoHu/Skill-Evo4GUI.
Review