WorkFactory maps intents into executable agent tasks

WorkFactory’s core insight is compelling: agent training data must be executable, internally supported, and repaired before it becomes useful supervision.

• –Information-support graphs connect professional intents to the evidence, rules, and files required to complete them
• –Probe executions detect missing or inconsistent task information before teacher trajectories are collected
• –WorkPilot models improved over matched Qwen baselines across four professional-work benchmarks
• –The 112,686-task corpus spans 612 occupations and 53 deliverable file types, offering unusually broad coverage
• –The pipeline is computationally expensive, averaging 11.656 million construction tokens per attempted task, making efficiency a key scaling constraint