Computer use
An agent operating a real GUI — screenshots in, mouse and keyboard actions out.
Computer-use agents perceive the screen (pixels and/or accessibility tree) and act through clicks, typing and scrolling, so they can drive any application without an API. OSWorld is the reference benchmark: 369 real tasks across Ubuntu, Windows and macOS. Best-agent success went from 12.2% at launch (2024) to about 66% in 2025 against a human baseline of 72.4%.