Abstract: As the critical link between the macro urban scale and the micro block scale, mesoscale urban design has the dual responsibility of coordinating two-dimensional functional order and three-dimensional morphological order. In the face of dynamically changing urban demands, conventional mesoscale design methods are constrained by static and singular function-morphology coupling and by relatively low adjustment efficiency. Drawing on reinforcement learning, this study develops an intelligent mesoscale urban design workflow comprising a network-control module, a scheme-iteration module, and a comparison-and-optimization module. Based on the spatial configuration of function-morphology coupling elements, the study identifies several interaction modes, including high coupling, core clustering, axis-driven development, function-led development, and morphology-led development, and applies them in an empirical design experiment. The results show that the proposed method substantially improves scheme diversity and design efficiency, strengthens adaptability, and offers an effective pathway toward dynamic interaction in future urban design.