Tag
This paper investigates execution bottlenecks in computer-use agents, comparing screen-only GUI-based approaches with skill-mediated CLI-based methods, identifying key performance differences.