The growing sophistication of contemporary smart grids, driven by the integration of large-scale and distributed renewable energy sources, two-way energy flows, and demand, requires sophisticated real-time stability analysis and prediction models. The conventional deep learning-based, single-agent, single-objective reinforcement learning frameworks do not apply to the dynamics of topology, decentralized decision-making, and the heterogeneity of interactions in large power networks. To overcome these issues, this paper proposes a Graph Transformer-based Multi-Agent Reinforcement Learning (GT-MARL) framework for real-time smart grid stability monitoring and anomaly detection. The model portrays the smart grid as a graph, with nodes representing generators, prosumers, loads, and sensors, and edges representing electrical and communication connections. Graph Transformer represents long-range dependencies and topological variations more effectively than traditional GNNs or LSTMs, thereby supporting spatiotemporal feature extraction under variable operating conditions. Based on this representation, a Multi-Agent Reinforcement Learning (MARL) framework is implemented, in which each agent learns independently to select optimal actions to improve local stability, reduce voltage/frequency variations, and achieve dynamic load balancing. Agents communicate via attention-based message passing over transformer embeddings, ensuring grid-scale consistency without a central authority. As well as, a combined anomaly prediction component detects cyberattacks, sensor failures, and irregular operating dynamics using transformer-based temporal attention. Simulations of real-time renewable-integrated testbeds show that the proposed GT-MARL model is more accurate, responds faster, maintains stability, and detects anomalies earlier than the baseline RL, GNN, and hybrid deep-learning models. The findings verify that the GT-MARL solution provides a scalable, smart, and sustainable solution to the next-generation autonomous smart grid activities. GT-MARL achieves voltage deviation 0.8%, frequency error 0.04 Hz, load imbalance 1.9%, anomaly accuracy 96.2%, detection latency 1.3s, communication overhead 1.1 MB per timestep.Keywords: Smart Grid Stability, Graph Transformer Networks, Multi-Agent Reinforcement Learning, Real-Time Anomaly Detection, Spatiotemporal Feature Extraction, Distributed Energy Resources (DERs), Renewable-Integrated Power Systems, Autonomous Grid Control