Enables multimodal models to operate computers using same inputs/outputs as human operators. Model views screen and decides series of mouse and keyboard actions to reach objectives. Autonomous computer control through visual understanding and action planning.