Cowork-Gym: Synthesizing Interactive Work Environments at Scale

Cowork-Gym. Synthesizing Interactive Work Environments at Scale. By Anni Zou and twenty others, from Alibaba Token Foundry. Published on the thirteenth of October. No code repository is linked to it, though it does publish its weights.

A cowork agent navigates dependencies across tools and files, while the user keeps changing what they want and editing the shared workspace underneath it. Building many environments like that is hard, because four things have to stay consistent with one another: the task objectives, the executable resources, the user's behaviour, and the criteria for success.

Cowork-Gym synthesizes these environments through a loop of build, execute, and revise. Execution evidence and review feedback are used to validate each generated environment and to repair it when it falls short.

To make the dynamics realistic, the user is simulated, and the simulator is stateful. It responds to what the agent does in two ways: through dialogue, and through direct edits to the shared workspace. With this, the authors synthesize sixty-four point nine thousand environments across fourteen application scenarios. They use them to post-train Qwen models at nine billion, twenty-seven billion, and thirty-five billion parameters, using supervised fine-tuning, reinforcement learning, and on-policy distillation.

Averaged across six agentic benchmarks, the nine billion model gains eight point seven points over its base checkpoint. The twenty-seven billion model gains six point four. The thirty-five billion model gains four point one. The twenty-seven billion model ranks second overall among the models evaluated, and posts the highest average scores on PawBench and ClawEval, ahead of models with over one hundred billion total parameters on both.

The gains shrink as the base model grows, yet they hold at every size, and the middle model ends up ahead of far larger ones on two benchmarks. Underneath the numbers, the bet is on consistency. An environment is only useful for training if its objective, its resources, its simulated user and its success check all agree, and the build, execute and revise loop is how Cowork-Gym keeps them agreeing at scale: environments that are checked by running them, rather than ones that merely look plausible on paper.