CoworkEnv synthesizes training environments for AI coworkers: Qwen3-VL-8B jumps from 46.6 to 64.1

How to train AI coworkers for real-world scenarios

How do you train AI coworkers to use tools, handle multi-step interactions, and adapt to changing workspaces? This is a central challenge in building agentic systems, and researchers have proposed a fundamentally new approach.

The team introduced CoworkEnv, a framework for synthesizing executable training environments. Unlike static datasets, CoworkEnv dynamically generates environments where models can learn to interact with tools and adapt to shifting conditions.

Architecture and components

The framework combines three key mechanisms:

• Capability-guided task generation — the system generates tasks tailored to the model's current abilities, ensuring progressive difficulty
• Programmatic validation — automatic correctness checks through code execution and result verification
• Feedback-based repair — a mechanism that automatically fixes flawed examples based on the feedback received

The system supports both major training approaches — supervised fine-tuning (SFT) and reinforcement learning (RL) — allowing teams to flexibly choose a strategy depending on the task.

Training results: compact models, powerful gains

The researchers trained a series of compact PocketWorker models using this framework. The most impressive result came from Qwen3-VL-8B-Instruct:

• Overall Claw-Eval score rose from 46.6 to 64.1 points (+17.5 points)
• On the multi-step task subset, the gain reached +31.9 points
• On the Claw-Eval Multi-turn block, the model climbed from 45.9 to 77.8 points

The experiments also used GDPO (gradient-guided DPO) for RL, which overall reached 57.2 points on Claw-Eval, trailing the fully trained SFT model but leaving room for further optimization.

Practical implementation details

The system includes a simple but effective grading-consistency gate — a check for logical consistency when scoring results. The rule works as follows: a complete and correct result gets the maximum score, an empty response gets zero, and any call to a prohibited tool zeroes out the safety score. This approach is so effective that the authors recommend it as a standard tool for other environment-synthesis pipelines.

The researchers are actively inviting feedback from the community and are open to discussing environment-synthesis methodology and training agents with long-horizon planning.