The growth of Internet of Things (IoT) ecosystems has led to continuous data generation across diverse environments, creating the need for efficient interaction among cloud, edge, and IoT layers. Existing coordination approaches face difficulties with unpredictable latency, workload imbalance, and resource waste. Traditional static policies cannot adapt well to dynamic traffic variations and device constraints. This work introduces a cross-layer machine learning model designed for predictive and adaptive coordination. The model combines graph neural networks (GNNs) with deep reinforcement learning (DRL) to capture spatial and temporal patterns of resource states, connectivity, and workload fluctuations. This integration supports a unified decision process across all layers, guiding task distribution and migration with adaptive policies rather than fixed rules. The proposed approach has been evaluated using simulations with varied IoT scales and workload scenarios. Performance has been measured using latency, throughput, task success rate, and energy consumption. Results demonstrate consistent gains, including faster response times, higher resource utilization, and reduced energy overhead compared with baseline scheduling methods. These findings confirm that the model provides a robust foundation for seamless cloud–edge–IoT interaction and supports reliable performance in dynamic environments.