Abstract: Intelligent agents realize autonomous operation to accomplish task objectives in the operational environment via the closed-loop mechanism consisting of perception, decision-making, action and feedback. Supported by Large Language Models (LLM), tool invocation, retrieval augmentation, operational observability and unified protocols accelerate the transition of intelligent agents from prototypes to commercial platforms and large-scale deployment. On the basis of the definition and connotation of LLM-based intelligent agents, this paper designed a general architecture for such agents and systematically reviewed the global technological evolution of intelligent agents, covering relevant technologies including tool-augmented LLM agents, embodied agents, multi-agent collaboration and enterprise-level platforms. Targeting the intelligent development requirements of China’s railway industry, the paper proposed typical application scenarios of intelligent agents focusing on intelligent maintenance, technical safety prevention and operational service, and clarified the form and core functions of each agent. Combined with the demands of engineering implementation, it put forward phased implementation routes and key development priorities to provide theoretical support for the high-quality intelligent transformation of the railway industry.