Abstract: Indoor fires pose a serious threat to lives and property. However, existing detection methods face issues such as high false alarm rates and a disconnect between early warning and emergency response. To address these challenges, this study proposes an intelligent fire monitoring agent system based on cloud-edge collaboration and multimodal large models, achieving full-process automation from fire perception to emergency decision-making. The system consists of a lightweight fire detection model and a cloud-based multimodal intelligent agent working in coordination. First, building upon the parameter-free feature extraction module HLFE, a lightweight detection model named HLLH-YOLO is developed. It reduces the number of parameters by more than 20% while maintaining detection accuracy comparable to models of the same level, enabling deployment on edge devices for real-time fire recognition. Second, a fire emergency intelligent agent named FireAgent, based on a vision-language large model, is constructed. It performs secondary verification of alarm-triggering images and automatically carries out fire situation analysis, alert notification, and generates response suggestions enhanced by Retrieval-Augmented Generation (RAG). Experiments show that the proposed detection model outperforms mainstream methods in both accuracy and lightweight design. Furthermore, the intelligent agent effectively addresses false alarm issues in fire detection, intelligently analyzes on-site fire conditions, and quickly generates emergency decision-making suggestions. This study provides a real-time, reliable, and integrated solution for indoor fire prevention and control.