As autonomous driving technology progresses towards large-scale commercialization, simulation testing has emerged as a crucial tool for validating its efficacy. However, the existing simulation methods are still insufficient in simulating the interaction of real complex behaviors, such as yielding and merging. To bridge this gap, this paper proposes an AI agent model based on Sim Transformer, which uses its powerful sequence learning ability to accurately model and predict traffic intelligent behavior. During model training, we integrated open-source trajectory data, simulation data, and roadside sensor data to learn realistic traffic behaviors, with the aim of enhancing the microscopic simulation capabilities of autonomous traffic systems. Experimental results demonstrate that this model exhibits significant advantages in simulating realistic vehicle interaction behaviors.