“Cursor's Composer 2.5 model is trained using targeted reinforcement learning with textual feedback to solve the credit assignment problem in long AI agent trajectories.”