The team introducing EmbodiedSWE says coding agents could autonomously solve complex robot tasks in its study, which covered precision, dexterity and long-horizon tasks. But it argues that running a coding agent for every robot action is expensive, slow and difficult to deploy safely. The team is exploring using successful solutions as teachers for fast robot policies, claiming that one solution can yield a large, diverse training dataset at low cost. Its proposed approach pairs a GPT-like reasoning “brain” that discovers solutions with a fast VLA or reinforcement-learning system for physical action.