GUI agents that click through apps and desktops can now patch their own playbooks on the fly, without retraining.
Researchers built EvoSkill-GUI, a training-free framework that treats agent "skills" as living documents instead of fixed scripts written once and left alone. Each skill ships as a multi-file package: executable plans, backup ways to locate on-screen elements, failure-recovery rules, accessibility utilities, and a record of past failures. The system runs on a reflect-revise-reuse loop - the agent revises its plan instantly when something breaks mid-task, then a separate critic reviews failed attempts in isolation and edits specific skill files through a restricted interface, rather than letting the agent rewrite anything it wants. The team tested it on three benchmarks spanning mobile and desktop: MobileWorld, AndroidWorld, and OSWorld.
This matters because GUI agents fail for mundane reasons: a pop-up appears, a button moves, a page loads slowly, and a plan drawn up before the run falls apart. EvoSkill-GUI improved multiple base models with no extra training, gaining up to 16.2% on MobileWorld, 6.0% on AndroidWorld, and 10.5% on OSWorld, and the evolved skill libraries kept helping on related tasks instead of needing to be rebuilt from scratch each time.
Those gains come from controlled benchmark environments, not the chaos of a real app pushing a surprise redesign overnight - the harder test is whether a self-patched skill file survives interfaces nobody built the benchmark to anticipate.