Abstract: With the rapid advancement of artificial intelligence, large language models (LLMs) have emerged as a promising technological paradigm for accelerating the intelligent transformation of the steel industry. Owing to their capabilities in semantic understanding, logical reasoning, knowledge organization, and cross-task generalization, LLMs offer a promising new pathway for addressing the limitations of traditional industrial artificial intelligence approaches, which are often restricted to isolated processes, narrowly defined tasks, and weak cross-process coordination. Against this background, the architectural construction and application logic of LLM-driven industrial intelligent agents for the steel industry have attracted increasing attention, particularly in relation to the development context, core capabilities, adaptation strategies, and representative application paths of industrial large models across the full steel production process. Starting from the evolution of mainstream LLM architectures and their industrial adaptation mechanisms, current progress shows how general-purpose models can be transformed into domain-oriented industrial large models through prompt engineering, retrieval-augmented generation, fine-tuning, and domain pretraining. In contrast to general LLMs designed for open-domain language tasks, industrial large models are fundamentally oriented toward constrained, verifiable, and operationally grounded decision support. On this basis, the system construction framework of steel-industry large models can be understood as a layered architecture composed of foundation models, domain enhancement modules, and scenario application modules. In particular, three major fusion modes for integrating process knowledge, industrial data, and model reasoning in steel production can be identified: serial fusion, parallel fusion, and embedded fusion, thereby clarifying the structural route through which LLMs may be incorporated into highly coupled metallurgical processes. Considering the characteristics of the steel industry, including dense multisource heterogeneous data, strong process constraints, complex operating conditions, and long decision chains, current research and industrial practice can be organized around three interrelated dimensions: data processing and dataset construction, fault diagnosis and root-cause analysis, and process scheduling optimization. In the first dimension, emphasis is placed on converting structured sensor streams, equipment logs, technical manuals, standards, and literature into domain-specific corpora and multimodal datasets suitable for supervised fine-tuning, reasoning tasks, and downstream evaluation. In the second dimension, LLMs—often combined with knowledge graphs, causal analysis, and multiagent reasoning—support the interpretable diagnosis of static defects, dynamic deviations, and systemic failures across the steelmaking, casting, and rolling processes. In the third dimension, LLM-related techniques help bridge perception, analysis, and execution for static and dynamic scheduling problems under multiple objectives, such as energy consumption, production rhythm, quality stability, and resource allocation. The role of LLMs in the steel industry has evolved from auxiliary analytical tools toward system-level intelligent decision-support hubs. Nevertheless, broader deployment still faces critical challenges, including multimodal data alignment under physical constraints, hallucination risks in high-stakes industrial settings, limited interpretability of model reasoning, and the difficulty of integrating multidisciplinary knowledge from metallurgy, materials science, control engineering, and artificial intelligence. Future development is expected to focus on high-quality industrial corpus construction, stronger physical consistency and explainability, multi-agent collaboration, industrial internet integration, and lightweight real-time deployment. Overall, this line of inquiry provides a structured reference for the systematic research, engineering implementation, and paradigm evolution of LLM-driven intelligent agents in the steel industry.