Next Point-of-Interest (POI) recommendation predicts a user’s next location based on their historical trajectory, but existing methods, such as collaborative filtering and neural network-based models, struggle with dynamic data and cold-start issues. This paper presents MAS4POI, a novel multi-agent system designed to improve the next POI recommendation by leveraging specialized agents. The system incorporates four agents: Manager, DataAgent, Analyst, and Reflector, each performing distinct tasks like coordinating system operations, preprocessing check-in data, generating initial recommendations and refining the outputs. MAS4POI achieves state-of-the-art performance on the NYC and TKY datasets, with Acc@1 of 77.30% and 76.43%, and MRR of 81.53% and 78.33%, respectively. Experiments also demonstrate that MAS4POI effectively mitigates the cold-start issue, with Acc@1 and MRR differences between inactive and very active users being only 0.92% and 3.2% on the NYC dataset, and 0.35% and 0.6% on the TKY dataset, respectively. Code available at: https://github.com/yuqian2003/MAS4POI