OpenComputer: Verifiable Software Worlds for Computer-Use Agents

Researchers introduced OpenComputer, a verifiable desktop-task framework for computer-use agents.

It combines app-specific state verifiers, self-improving verification, synthetic task generation, and trajectory-based evaluation across 33 desktop applications.

The framework tracks success through observable application state, making results more auditable than LLM-as-judge methods.

OpenComputer aligns better with human judgment and reveals major performance gaps in both frontier and open-source models.