Computer-Use 2.0: Agents Just Got Multi-Cursor — Francesco Bonacci, Cua

Computer-using agents are evolving from screen-takeover models to background-driven automation. The CUA driver enables AI agents to interact with operating systems—macOS, Windows, and Linux—using accessibility trees and pixel-level observation without disrupting the user. Evaluating these agents requires rigorous benchmarks like CUABench, which currently tracks over 130 tasks across five platforms; however, current model performance remains limited, with success rates dropping significantly on complex tasks like blank schematic creation. To optimize the high costs associated with reinforcement learning, infrastructure must shift toward demand-based autoscaling for sandboxes. By maintaining a dynamic pool of environments, developers can eliminate GPU idle time and ensure efficient, scalable training cycles. This integrated approach—combining background-capable drivers, standardized evaluation frameworks, and optimized infrastructure—is essential for advancing agent reliability and performance in real-world desktop environments.