Brilliant new paper from the Qwen team. It provides insights into where agent training environments actually come from. Terminal agent trajectories have accumulated at scale while realistic executable environments are scarce. Environments are what post-training needs, since
