Agent-Based Network Architecture and Resource Management for Vehicle-Road-Cloud Collaboration

Large artificial intelligence models (LAMs) have introduced new opportunities to 6G networks, whereas their integration into vehicle-road-cloud collaborative systems remains challenging due to insufficient model coordination and difficulties in dynamic resource adaptation. To address these issues, this article proposes an agent-based network architecture and a resource management method for vehicle-road-cloud collaboration. In this architecture, the cloud agent conducts largescale pre-training to characterize global resource allocation dynamics, while lightweight models are deployed at the road-side unit (RSU) agent and vehicle agent to enable real-time inference and decision-making. To balance model performance and resource consumption, we introduce a lifecycle management strategy for the agent network. By analyzing the relationship between performance gain and the consumption of communication and computation resources, we jointly optimize training data sampling rates, model pruning ratio, and resource allocation strategies. This enables dynamic adjustment of the update frequency for cloud-based large models and the fine-tuning frequency for edge-side models, thereby ensuring continuous and high-efficiency model inference in resource-constrained dynamic scenarios. Additionally, we design a decision LAM (D-LAM) for single-task scenarios, which employs sequence modeling and temporal embedding to achieve dynamic coordination of multi-dimensional resources. We further develop a decision multi-task LAM (DM-LAM) to address complex multi-task scenarios, which integrates a mixture-of-experts (MoE) mechanism with a gating network to support both personalized resource allocation for individual tasks and global optimization across heterogeneous tasks. Experimental results demonstrate that the proposed method significantly improves resource utilization efficiency and achieves 10 $1.2% higher task completion rates compared to decision transformer model.