Researchers have released OpenGCUA, an open-source project providing a full stack for generalist computer-use agents that observe, act, and organize work across both visual and programmatic interfaces. The release includes the implementation, prompts, and evaluation trajectories from eight models, along with early versions of a real-device harness and a cross-device SDK. Core techniques include using code execution to guide intermediate actions and shifting certain observation and task-state decisions from fixed harness rules to the model itself. On the OSWorld-v2 benchmark, the full configuration raised mean partial task scores by 19.6 points and cut steps per task by 53% versus batched GUI-only execution.
