Cua's Computer-Use 2.0 account explains how to judge an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → by the work it leaves behind. Cua-Bench separates a task into setup, the agent's attempt and an evaluator that inspects the resulting files or application state. In its KiCad example, the evaluator compares saved circuit components and connections with a reference.
This supports the distinction between finished work and partial progress in Cua's electronics-task comparison. The company also describes testing evaluators with an agent that tries to earn credit while violating the task. That checks whether the scoring rule rewards the intended result or can be fooled by a shortcut.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
