BehaviorGPT is a novel decoder-only autoregressive architecture designed to simulate the sequential motion of multiple traffic agents in autonomous driving systems. Unlike traditional encoder-decoder structures that manually split historical and future trajectories, BehaviorGPT treats each time step as the "current" one, simplifying the model architecture and improving data utilization. The Next-Patch Prediction Paradigm (NP3) enables the model to reason at the patch level of trajectories, capturing long-range spatial-temporal interactions. This approach maximizes the utilization of agent data and avoids wasting any learning signals available in the time series. BehaviorGPT was evaluated on the Waymo Open Sim Agents Challenge, achieving top rankings across several metrics, including a realism score of 0.741 and a minADE metric of 1.540, with a significant reduction in model parameters (91.6% less compared to state-of-the-art models). The method demonstrates superior performance in multi-agent and agent-map interactions, making it a robust and efficient choice for traffic agent simulation in autonomous driving research.BehaviorGPT is a novel decoder-only autoregressive architecture designed to simulate the sequential motion of multiple traffic agents in autonomous driving systems. Unlike traditional encoder-decoder structures that manually split historical and future trajectories, BehaviorGPT treats each time step as the "current" one, simplifying the model architecture and improving data utilization. The Next-Patch Prediction Paradigm (NP3) enables the model to reason at the patch level of trajectories, capturing long-range spatial-temporal interactions. This approach maximizes the utilization of agent data and avoids wasting any learning signals available in the time series. BehaviorGPT was evaluated on the Waymo Open Sim Agents Challenge, achieving top rankings across several metrics, including a realism score of 0.741 and a minADE metric of 1.540, with a significant reduction in model parameters (91.6% less compared to state-of-the-art models). The method demonstrates superior performance in multi-agent and agent-map interactions, making it a robust and efficient choice for traffic agent simulation in autonomous driving research.