Spatio-temporal forecasting presents significant challenges across domains from urban traffic analysis to epidemiological modeling. Despite advances in graph neural networks (GNNs), current approaches often lack the necessary adaptability for diverse real-world applications. We introduce AGNet, a specialized encoder-decoder architecture that addresses fundamental limitations in existing spatio-temporal models. Our architecture combines a configurable GNN encoder that effectively captures spatial dependencies with a GRU decoder for temporal sequence modeling, complemented by an MLP for accurate prediction generation. The key innovation of AGNet lies in its unified preprocessing framework that transforms data into standardized graph representations through nodes, edge indices, and attribute vectors. Extensive experiments on benchmark datasets including PeMSD7, METR-LA, and Chickenpox demonstrate that AGNet consistently outperforms state-of-the-art methods, achieving substantial improvements in both RMSE and MAE metrics. This work introduces a practical and well-validated framework that advances the current benchmarks and supports broader use in spatio-temporal forecasting.