Cowork-Gym Scales Training for Adaptive Work Agents

Introduction

Reliable work agents need more than the ability to call tools. They must track dependencies across files and applications, interpret changing requirements, ask for clarification when necessary, and revise work after a user edits a shared workspace. Many existing training tasks are static: the instruction, resources, and success condition are fixed in advance. That makes it difficult for an agent to learn the iterative nature of real collaboration.

Cowork-Gym addresses this problem by synthesizing executable environments for human–agent work at scale.

A build–execute–revise pipeline

The framework is organized around a build–execute–revise loop. It first creates an environment from task objectives and available resources. The generated task is then executed, allowing the system to collect evidence about whether its tools, files, dependencies, and success criteria actually work together. Review feedback is subsequently used to repair inconsistencies and improve the environment.

This design makes executability part of environment generation rather than an afterthought. Running a task can reveal mismatched acceptance criteria, unavailable resources, or missing dependencies that may be difficult to detect through text inspection alone. The supplied material does not report error distributions or repair success rates, so the contribution should be viewed primarily as a production and validation framework rather than a claim about one specific task solver.

Cowork-Gym also includes a stateful user simulator. Instead of issuing one instruction and disappearing, the simulated user reacts to the agent’s actions through dialogue and can directly edit the shared workspace. This creates training opportunities around requirement clarification, plan changes, and revision after feedback.

Scale and reported results

The authors synthesize 64,888 environments covering 14 application scenarios and 17 capability labels. They use these environments for post-training Qwen models at 9B, 27B, and 35B-A3B scales, applying supervised fine-tuning, reinforcement learning, and on-policy distillation.

Across six agentic benchmarks, average scores improve over the corresponding base checkpoints by 8.7, 6.4, and 4.1 points, respectively. The 27B model ranks second overall among the models evaluated in the study. It also records the highest average scores on PawBench and ClawEval, exceeding evaluated models with more than 100B total parameters on both benchmarks.

These are results under the authors’ reported evaluation setup and should not automatically be interpreted as universal superiority in every workplace scenario. The supplied material does not include per-benchmark scores, environment difficulty statistics, or detailed ablations. Those details will matter for assessing generalization.

Why it matters

Cowork-Gym shifts agent training from simply completing a fixed instruction toward maintaining progress in a changing workspace. If environment generation, execution checks, and user feedback can be combined reliably, training data may cover more long-horizon collaboration patterns without requiring every scenario to be written manually.

The approach also has clear open questions. A simulated user may not capture the preferences, communication styles, or editing habits of real people. Generated environments may overrepresent familiar patterns, and benchmark gains may not directly translate into workplace productivity. Still, the framework points to an important direction: training environments themselves should be executable, inspectable, and continuously revised rather than treated as static prompts.

Source: Hugging Face Daily Papers