benchmarks

Synthetic GUI Benchmarks Fail When the App Doesn't Actually Run

Training computer-use agents currently runs into a clean wall: native desktop applications are painful to spin up at scale, and browser toys lack workflow depth.