Mac computer-use agent that paired a small grounding VLM with GPT-5's reasoning over screen captures. It clicked, typed, and ran through RPA workflows; the interesting part was using the CLI directly whenever possible instead of pretending every computer action had to be pixel-level grounding.
I also used it to test how far strong computer-use models could get one-shot on games. Claude Opus could play League of Legends badly but coherently from a plain prompt.