EnvCraft Synthesizes Executable Environments for Claw-Like Agent RL

Summary

EnvCraft is an automated framework designed to address the shortage of interactive environments for training autonomous, Claw-like agents with Agentic Reinforcement Learning. The authors argue that existing synthetic environments mainly expose tool-calling endpoints and do not support the stateful, end-to-end workspaces these agents must operate in. EnvCraft combines an environment synthesis engine, which creates sandbox-isolated workspaces, with a topology-aware data generation engine that produces coherent task trajectories. The system synthesizes 139 interactive environments and approximately 20,000 complex tasks for training. Experiments with Qwen3 and Qwen3.5 models ranging from 8B to 32B report gains of up to 11.9% on Claw-style benchmarks and up to 8.0% on general tool-use benchmarks. The method also reduces inference token cost. The reported results suggest that executable synthetic environments can provide robust and generalizable learning signals for Agentic RL training.