Multi-agent system technology has been widely used in power system regulation, warehousing and logistics, etc. However, the accurate evaluation and deep optimization of the performance of such systems is still a frontier issue to be solved, and it is still mainly in the experimental and exploration stage. In this paper, a new attribute-based multi-agent design method is proposed, which is combined with performance analysis to carry out performance evaluation in the system design process. The design method takes the necessary attributes of the system as the core, adopts the top-down top-level design logic, starts from the functional requirements at the macro level, and refines step by step to the specific realization at the micro level, so as to ensure that the system design is closely developed around the actual needs. We have incorporated the “Off-Policy” evaluation method into this system and integrated it into the model verification step of the system design step. This evaluation method has unique advantages in the field of reinforcement learning, which can evaluate and compare the potential effects of other strategies while keeping the current operational strategies unchanged, and provide strong support for coping with the complexity and dynamic challenges in multi-agent systems. During the design process, we will carry out continuous evaluation and verification of the system: if the system can meet the preset attribute requirements, the design will enter the subsequent implementation phase to continue to promote; Conversely, if the system fails to meet the preset attribute criteria, an iterative optimization process is initiated to continuously adjust and refine the design until all attribute requirements are met.