In this paper, we propose a multiagent collaboration decision-making method for adaptive intersection complexity based on hierarchical reinforcement learning—H-CommNet, which uses a two-level structure for collaboration: the upper-level policy network fuses information from all agents and learns how to set a subtask for each agent, and the lower-level policy network relies on the local observation of the agent to control the action targets of the agents from each subtask in the upper layer. H-CommNet allows multiagents to complete collaboration on different time scales, and the scale is controllable. It also uses the computational intelligence of invehicle intelligence and edge nodes to achieve joint optimization of computing resources and communication resources. Through the simulation experiments in the intersection environment without traffic lights, the experimental results show that H-CommNet can achieve better results than baseline in different complexity scenarios when using as few resources as possible, and the scalability, flexibility, and control effects have been improved.