CityPlanner uses sandbox agents for urban optimization

CityPlanner is a sandbox-agent framework that uses an executable environment to perform urban planning tasks. It employs atomic-task reinforcement learning to decompose complex trajectories into initial construction and feedback-based refinement phases.