WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training

WorkFactory. Turning Professional Intents into Executable Work for Scalable Agent Training. By Ziyang Zeng and twelve others, from IQuest Research, Zhejiang University, and four other institutions. Published on the twenty-seventh of October.

Training agents to do professional work by supervision needs a large supply of diverse professional tasks, each one able to yield a reliable expert demonstration. The weak point is the task itself. When information the task depends on is missing or inconsistent, the teacher model trying to execute it gets obstructed, and the demonstration it leaves behind is weaker supervision than it should be.

WorkFactory is a scalable framework for constructing professional tasks that are coherent and executable before any teacher attempts them. It starts from a professional intent, meaning the work someone wants done, and plans at the level of information: what information the intended outputs require, how those pieces depend on each other, and where each one is located.

That plan takes the form of an information support graph, which then guides how the task is instantiated. The graph is an explicit record of the dependencies and locations of the information the outputs need.

Planning alone is not trusted. Probe executions are run against the instantiated task to identify support failures, places where the needed information is not actually supported, and those failures guide repairs. The repairs are constrained to preserve the intended work, so a fix makes the task executable without turning it into a different task.

The framework produced one hundred and twelve thousand six hundred and eighty-six tasks, spanning six hundred and twelve occupations. Teacher trajectories collected on them were used to fine-tune Qwen models at three scales: four B, twenty-seven B, and thirty-five B A three B. The result is the WorkPilot family.

All three models improve over their own base models across four professional-work benchmarks. The largest gain quoted is on GDPval-AA v2.1, where the biggest WorkPilot model rises from eight hundred and seventy-nine Elo to one thousand five hundred and sixty-five Elo.

Scaling analyses show diminishing returns that depend on the model, with execution length first increasing and then decreasing. Controlled component analyses isolate the two stages: planning and repair each reduce the task environment defects that probes diagnose, and they improve downstream performance.

The bet here is that the quality of agent supervision is set upstream, in the task, before the teacher ever acts. WorkFactory treats a task as something to be engineered and checked: plan what information it needs, probe whether that information is really there, and repair what is missing without changing the job being asked for.