WorkFactory Turns Professional Intent into Executable Agent Training Tasks

Introduction

Training an agent for professional work requires more than teaching it to reason or call tools. The task environment must also provide the right documents, facts, dependencies, and access paths. In a real workflow, a missing attachment or inaccessible web page can prevent an otherwise capable model from completing the assignment. If such failures are used directly to collect teacher trajectories, the resulting supervision may reflect broken environments rather than useful work skills.

WorkFactory, presented in a paper listed by Hugging Face Daily Papers, treats task construction as a core part of working-agent training. The framework starts with a professional intent and builds a task that can be executed coherently at scale.

From intent to an information-supported task

The first stage is information-level planning. Instead of generating only a natural-language task description, WorkFactory constructs an information support graph around the intended output. The graph represents what information is required, how those pieces depend on one another, and where the information should be available. This gives the generated task an explicit relationship between its goal, evidence, and execution environment.

The system then runs a probe execution. Using only the inputs visible to the solver, the probe attempts to identify missing support or other blockers. When a problem is found, the framework repairs the blocking gap while preserving the original work intent as much as possible. A teacher model can then execute the repaired task and produce a demonstration trajectory.

This probe-then-repair loop is important because it puts environment quality checks before expensive trajectory collection. The paper’s controlled analyses report that both planning and repair reduce defects detected during probing and improve downstream performance.

Scale and model results

WorkFactory produces 112,686 tasks spanning 612 occupations. The collected teacher trajectories are used to fine-tune WorkPilot models at 4B, 27B, and 35B-A3B scales. All three versions improve over their respective base models on four professional-work benchmarks. The largest reported jump is for WorkPilot-35B-A3B on GDPval-AA v2.1, where the score rises from 879 to 1565 Elo.

These results suggest that the quality of a work-agent dataset depends on more than occupational diversity or raw task count. A task must also be executable enough for its demonstration to represent the intended capability.

Scaling brings efficiency as well as diminishing returns

The scaling analysis shows that more data does not produce linear gains. Early additions can deliver substantial improvements, while later additions provide smaller benefits as the student approaches the capability ceiling represented by the teacher. At the same time, execution length first increases and later decreases in the reported experiments.

A plausible interpretation is that additional data helps the model learn not only what to do, but also how to plan with fewer unnecessary steps and tool calls. However, shorter trajectories should not automatically be treated as better deliverables; the supplied material does not establish that execution efficiency and output quality always move together.

Why it matters

WorkFactory reframes synthetic data generation as environment engineering. The central question is not simply whether a task sounds professional, but whether its required information and tools support a successful execution. That perspective is useful for researchers building general-purpose agents across many occupations.

The approach also inherits limitations from external tools. The authors note that inaccessible websites account for many web-fetch failures, while web search fails less often. They identify rejection sampling, harder tasks, and reinforcement-learning scaling as possible ways to move beyond the current teacher and environment ceilings.

Source: Hugging Face Daily Papers