The transition to a decentralized power grid, driven by the high penetration of Distributed Energy Resources (DERs) and massive IoT deployments, necessitates ultra- reliable, low-latency communication for real-time control and stability. Time-critical applications, particularly Fault Detection, Isolation, and Restoration (FDIR) and Wide-Area Monitoring Systems (WAMS), are fundamentally compromised by variable network latency. Existing solutions, including static-edge computing architectures, fail to provide adaptive performance, as they are typically oblivious to the dynamic state of both the physical grid and the communication network. This paper introduces the Adaptive Latency Reduction Smart Grid Architecture (ALR-SGA), a novel distributed intelligence framework. ALR-SGA leverages a two-part mechanism: (1) a Graph Neural Network (GNN) model that performs topology-aware prediction of network congestion by fusing traffic data with physical grid-state information, and (2) a distributed Deep Reinforcement Learning (DRL) agent at each edge node that makes autonomous data routing and processing decisions. The DRL agent's reward function is uniquely tied to grid stability metrics, enabling it to prioritize data streams based on their real-time impact on grid operations. Through co-simulation of the IEEE 39-bus system (MATPOWER) and a detailed communication network (ns-3), we demonstrate that ALR-SGA significantly outperforms benchmark models. Results show a reduction in end-to-end latency for critical FDIR messages by over 35% during high-congestion fault scenarios, providing a viable architecture for next-generation, resilient grid control.