Abstract: Indoor fires severely threaten human lives and property. Although deep learning-based fire detection has advanced significantly, existing vision-based methods present two critical limitations: a high false alarm rate caused by environmental interference and lighting variations; and a disconnect between early warning and emergency response, where most systems merely trigger alarms without providing actionable decision support. Deploying computationally intensive detection models on resource-constrained edge devices remains challenging, while transmitting all video data to the cloud incurs excessive bandwidth and latency costs. Hence, this study proposes a cloud-edge collaborative intelligent agent system for indoor fire monitoring, which affords full-process automation from fire perception to emergency decision-making. The system integrates a lightweight detection model at the edge and a multimodal intelligent agent in the cloud, thus forming a cascaded “edge screening–cloud deliberation” pipeline. At the edge layer, a novel parameter-free feature-extraction module named Hodges–Lehmann-based feature extraction (HLFE) is proposed and embedded into the You Only Look Once verion 8 (YOLOv8s) architecture to construct an HLLH-YOLO model. Inspired by robust statistics, HLFE extracts three complementary statistical features—maximum pooling, adaptive average pooling, and the L2 norm—from feature maps and fuses them via pairwise averaging followed by a median operation. This design effectively suppresses outliers and noise-induced deviations without introducing any additional trainable parameters. Combined with varifocal loss for jointly optimizing classification and localization, HLLH-YOLO maintains detection accuracy comparable to mainstream models while reducing parameters by over 20%, thus enabling real-time deployment on edge devices for continuous video stream analysis. Once HLLH-YOLO triggers an alarm, the suspicious frame is uploaded to the cloud-based FireAgent. Instead of directly invoking a generic vision-language model (VLM), this study first constructs a dedicated multimodal evaluation dataset containing both actual fire scenarios and challenging interference samples. Subsequently, several open-source VLMs are benchmarked to identify Qwen2.5-VL-7B as the optimal backbone for accurate fire-scene understanding and secondary verification. FireAgent uses this model to rigorously verify the alert, thus eliminating residual false positives from the edge stage. Upon confirmation of a genuine fire, the visual information is automatically converted into a structured textual disaster description. Next, a large language model processes this description with retrieval-augmented generation. Consequently, relevant historical fire cases and safety protocols are retrieved to enable context-aware risk analyses, alarm notifications, and actionable emergency response suggestions, thereby closing the loop from perception to decision. To validate the framework, a dual-level dataset is constructed: a large-scale image dataset comprising approximately 30000 images encompassing diverse indoor scenarios, lighting conditions, and smoke occlusion levels for training and evaluating detection models, and a customized multimodal reasoning dataset for assessing VLM inference capabilities. Comparative experiments demonstrate that HLLH-YOLO outperforms mainstream methods in balancing accuracy and lightweight design, while FireAgent effectively resolves false alarms and rapidly generates reliable emergency decisions. This study offers a holistic cloud-edge solution that synergizes a structurally improved parameter-free detection module with a multimodal agent. By addressing both the algorithmic constraints of edge detection and the limitations of emergency response, the proposed system provides a real-time, reliable, and integrated approach to indoor fire prevention, monitoring, and emergency management.