Large Language Models have swiftly moved from passive language interfaces to autonomous claw‑like agents that can carry out long‑horizon tasks within stateful workspaces. Agentic Reinforcement Learning offers a promising route to optimize such agents, yet its scalability is severely limited by the scarcity of interactive training environments. Existing synthetic environments are confined to tool‑calling endpoints and fail to meet the end‑to‑end real‑world requirements of claw‑style agents.\ \ To bridge this gap, we introduce EnvCraft, an automated framework that synthesizes executable environments and generates scalable training data. EnvCraft consists of two main engines:\
- Environment Synthesis Engine: builds sandbox‑isolated workspaces, ensuring safe and repeatable interactions;\
- Topology‑Aware Data Generation Engine: creates coherent task trajectories based on the structural topology of each environment, preserving logical consistency across tasks.\ \ Using this framework, we synthesize 139 interactive environments comprising roughly 20 000 complex tasks for Agentic RL training. Experiments on Qwen3/3.5 models (8B‑32B) demonstrate up to +11.9% improvement on claw‑style benchmarks and +8.0% on general tool‑use benchmarks, while also reducing inference token cost.\ \ These findings confirm that synthesized executable environments provide robust and generalizable learning signals, laying a solid foundation for large‑scale Agentic RL training.\ \ Review