Abstract: To address the inaccuracy in calculating transfer times for urban rail transit, this paper proposes a multi-agent simulation method that takes passenger behavior and dynamic congestion into account. A microscopic simulation model is constructed to integrate passenger path choice behavior with bi-directional congestion effects. By assigning independent decision-making capabilities and speed attributes to each passenger agent, the entire transfer process from alighting to boarding is simulated. Through a case study of Beijing's Chaoyangmen Station, the model's ability to reproduce peak-hour congestion is demonstrated, with a deviation of 4.28% from the actual walking time. Bi-directional flow conflict was identified as the key factor for increased transfer time and reduced walking speed. Notably, when the bi-directional congestion coefficient was increased from low to a high level, the total walking time surged by 72.1%. By detailing microscopic behaviors, prediction accuracy is enhanced, and scientific decision support for passenger flow optimization and train schedule coordination is provided.