The Dynamic Feeder Vehicle Routing Problem (DFVRP) is a complex variant of the Vehicle Routing Problem (VRP), which is designed to address dynamic demands and multi-vehicle collaborative scheduling in urban logistics distribution. However, traditional algorithms are often hindered by slow response times and low coordination efficiency among multiple vehicles in dynamic environments. To improve efficiency and solution quality in dynamic path planning, a multi-agent deep Q-learning (DQN) search algorithm incorporating ant colony algorithm concepts is proposed in this paper. The algorithm consists of three agents: Truck agent, Motorcycle agent, and Adjustment agent. Among them, the Truck agent and the Motorcycle agent are responsible for selecting actions to serve customers through the DQN network. The ACO strategy is incorporated into their actions to enhance decision-making accuracy, while network parameters are optimized through a reward mechanism. Route generation is achieved through the collaboration of these two agents. Meanwhile, the Adjustment agent is responsible for selecting optimization actions based on route states, where a reward function is constructed using target value changes and route similarity to refine network performance. When dynamic customer requests are triggered, a multi-agent coordination mechanism is activated for route replanning. Notably, decisions are made by the agents based on learned policies, through which high-quality actions in the new state are rewarded, enabling the rapid generation of optimized route sequences. Experimental results have been obtained, which indicate that significant improvements in both route quality and computational efficiency are achieved by the proposed MDQN compared to baseline methods.