CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents

CoworkEnv. Synthesizing Training Environments for Long-Horizon Cowork Agents. By Zhongwei Xie and twelve others, from the Hong Kong University of Science and Technology, the Chinese University of Hong Kong, and two other institutions. Published on the eighteenth of October. Its code repository is worked out: about fifty-five source files, with a readme and a licence.

An AI coworker has to coordinate tools, communicate over many turns, and act on a workspace whose state keeps changing, all while respecting the constraints of a workflow. Training those capabilities needs experience to train on, and building that experience is a job in itself. It takes coherent requests, service records, tool interfaces, and checks that can grade what the agent did.

CoworkEnv is a framework for synthesizing that experience as executable training environments. It is capability-guided: the requirements of a workflow steer how each task is constructed.

What gets generated is then checked programmatically. Validation, together with a feedback repair step, examines the generated specifications and how the grading behaves. Once the environments exist, they feed two kinds of training. Runs by a teacher model yield curated demonstrations for supervised fine-tuning. Interactions by the student supply rollouts and rewards for reinforcement learning. With these resources the authors train a family of compact models called PocketWorker.

On Qwen3-VL-8B-Instruct, supervised fine-tuning alone raises the mean score on Claw-Eval from forty-six point six to sixty-four point one. Within that, the Multi-turn tasks gain thirty-one point nine points. A second benchmark, ClawMark, gets a fine-tuning-only extension, which shows the approach carries across coworker benchmarks.

For reinforcement learning, the authors compare training dynamics across configurations that differ in how rewards are processed and how credit is assigned. Training rewards improve, which they offer as preliminary evidence that the environments support policy optimization.

The firm result is the fine-tuning one, and the largest single piece of it lands on multi-turn work, the kind of communication across turns that the framework set out to train. The reinforcement learning side is framed modestly, as a sign that these environments can carry that kind of training too.