The rapid increase in large-scale visual data in cloud computing systems appeals for intelligent, autonomous and highly scalable AI- driven image classification systems. This work proposes a novelty of agentic-based DL framework applied to image classification in cloud computing system and increases accuracy through enhanced operational efficiency with a new agent-driven AI system. The DL architecture is adaptable and closely connected to an autonomous AI agent. The new proposed architecture includes an agent driven by hybrid attention guided Convolutional Neural Network (CNN) that adapts during pre-processing and tuning of model to real-time feedback and performance that is capable of controlling image pre-processing inference procedures. With deployment on a cloud-native architecture, the proposed framework increased capability for distributed training, elastic resource allocations and large-scale inferences. A series of experiments conducted on standard image datasets proved the proposed agentic DL framework outperformed the existing methods, this validates the effectiveness of agentic intelligence in the development of next generation intelligent and autonomous cloud-based image classification systems.