CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents
Abstract
CoworkEnv synthesizes executable training environments for AI coworkers using workflow-guided task construction and programmatic validation, enabling supervised fine-tuning and reinforcement learning of compact PocketWorker models that improve multi-turn coordination and tool use.
AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints. Constructing training experience for these capabilities requires coherent requests, service records, tool interfaces, and grading checks. We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers. Workflow requirements guide task construction, and programmatic validation with feedback repair checks generated specifications and grading behavior. Teacher executions yield curated demonstrations for supervised fine-tuning (SFT), while student interactions provide rollouts and rewards for reinforcement learning (RL). Using these resources, we train compact PocketWorker models. For Qwen3-VL-8B-Instruct, SFT raises the mean score from 46.6 to 64.1 on Claw-Eval, including a 31.9-point gain on Multi-turn tasks. An SFT-only extension to ClawMark demonstrates applicability across coworker benchmarks. We further compare RL training dynamics across configurations that vary reward processing and credit assignment. Improving training rewards provide preliminary evidence that the constructed environments support policy optimization.
Community
How can we build training environments for AI coworkers that need to use tools, handle multi-turn interactions, and operate in changing workspaces?
We introduce CoworkEnv, a framework that synthesizes executable environments for training these agents. It combines capability-guided task generation with programmatic validation and feedback-based repair, supporting both supervised fine-tuning and reinforcement learning.
Using CoworkEnv, we train compact PocketWorker models. On Claw-Eval, SFT improves Qwen3-VL-8B-Instruct from 46.6 to 64.1 (+17.5 points), with a +31.9-point gain on multi-turn tasks.
We welcome feedback and discussion on environment synthesis and training long-horizon agents!
Great to see such a strong multi-turn result for a compact model: SFT taking Qwen3-VL-8B from 45.9 to 77.8 on Claw-Eval Multi-turn is a big jump. The grading-consistency gate (a complete record must get full marks, an empty one zero, and a single prohibited call zeroes the safety score) is a simple, cheap check that more environment-synthesis pipelines should copy. One thing I couldn't place in Table 3: the RL runs start from a shared supervised warm-up checkpoint, and GDPO ends at 57.2 Overall versus 64.1 for the full-data SFT model. Could you share the warm-up checkpoint's own Claw-Eval score, so the RL contribution on top of it can be read directly?
This is an automated message from the ResearchStudio team.
We created an interactive ResearchStudio Reel for this paper. It includes a visual poster, a video, and a blog, all available for download in editable formats.
Open the ResearchStudio Reel →
Download all files from Hugging Face
Please give this comment a thumbs up if you find the Reel helpful!
Want to explore or create Reels for more papers? Visit the ResearchStudio demo.
Get this paper in your agent:
hf papers read 2610.24960 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper