A computer-use agent for Mac that reasons, plans, and executes clicks and keystrokes. Trained a small grounding VLM paired with GPT-5's reasoning and superimposed it on screen captures. Tested whether computer-use models can play games one-shot — gave Claude Opus a simple prompt like "play League of Legends" and it figured out the controls.
Built for a government entity to automate RPA workflows with pure agents. The paradigm shift is using the computer's CLI directly rather than pixel-level grounding.