CAD Teaches Agents Real Workflows
Professional computer-use is gaining the data and evaluation infrastructure needed for repeatable learning. A new open dataset contains more than 1,000 hours of synchronized CAD screens, actions, files and rubrics; CADBench measures agents on 105 realistic Fusion 360 tasks and finds a best pass rate of only 24.6%; and Microsoft's skill recorder reconstructs observed desktop work into reusable skills. Together they turn expert software workflows into trainable, auditable artifacts rather than one-off demonstrations.