Computer-Use Agents
Legacy systems with no API finally get an automation path. This covers how computer-use agents read the screen, locate controls and click, and why permission prompts and sandboxes have to ship before the capability does.
Many business systems never exposed an API, and computer-use agents sidestep it: read the screen, locate controls, then click and type. Web testing and repeated entry now fall to a model walking the interface. Open source moves fastest: MAI-UI maps instructions to coordinates, GELab-Zero drives Android with a 4B model, UI-TARS-2 holds long tasks together. A wrong click costs far more than a wrong answer, so approvals and sandboxes come first.
