Abstract: To address the problems of rigid diagnostic models, difficult cross-scenario migration, and the disconnection between algorithmic models and operation-and-maintenance knowledge in health monitoring of data center chiller units, this paper proposes an AI Agent-based health monitoring method. Inspired by the collaborative logic of preliminary screening and comprehensive consultation in human medical diagnosis, the proposed method constructs a diagnostic framework that integrates univariate threshold monitoring with multivariate kernel principal component analysis (KPCA). Based on the DeepSeek-V3 large language model, the AI Agent realizes automatic planning of monitoring strategies, invocation of algorithmic tools, and generation of executable code. Experiments on the public ASHRAE RP-1043 dataset show that the proposed AI Agent can accurately reproduce the diagnostic logic for seven typical faults, including reduced condenser water flow, reduced chilled water flow, and refrigerant leakage. Through a standardized function-calling mechanism, the method improves the migration efficiency and deployment flexibility of diagnostic models across different chiller units. The results further demonstrate that, with the collaboration of univariate threshold monitoring and multivariate KPCA tools, the proposed method can effectively reduce the false alarm rate in multi-condition and multi-unit health monitoring. Meanwhile, while maintaining diagnostic accuracy, the monitoring code generated by the AI Agent reduces the number of lines of code by approximately 8.6% compared with manually written code. This study provides a scalable and easy-to-deploy automated solution for fault prediction and health management of critical data center infrastructure.