CUA-SWE: When Computer-Use Agents Meet Visual Software Engineering

The authors introduce CUA-SWE, a benchmark, environment, and evaluation pipeline for software engineering with computer use. CUA-SWE addresses the gap in studying integrated development processes where agents must connect visual observations from graphical interfaces with the responsible code, then interact with the application to verify repairs. Spanning four software engineering domains, the benchmark requires agents to modify code and configuration, execute commands, interact with running software, and inspect visual feedback within the same task. Each task includes deterministic tests that verify whether the resulting software satisfies requirements.

CUA-SWE provides a unified testbed spanning four software engineering domains.

The benchmark integrates executable correctness criteria for the resulting software.