Agents trained on hybrid GUI+CLI tasks become significantly more efficient at real computer work, reducing steps by 37-57% and tokens by 30-60% compared to GUI-only approaches.
CUA-Universe is a system that turns real desktop applications into training environments where AI agents learn to work like humans—combining visual interface clicks with command-line commands. The system automatically adapts applications, generates realistic tasks, and trains agents to switch between GUI and CLI efficiently, resulting in agents that complete tasks faster and with fewer mistakes.