Topology awareness and scalable, adaptive network control have become critical with the development of 5G/6G, the Internet-of-Things, vehicular networks, and edge computing. Traditional rule-based and centralized networking models are unable to support dynamic topologies, heterogeneous traffic models, and demands with strict quality-of-service requirements. Structural and topological dependencies are encoded using graph neural networks (GNNs) and combined with deep reinforcement learning (DRL) to make decisions sequentially. Exploiting rewards is an avenue toward intelligent end-to-end network optimization. This review is a systematic examination of modern GNN–DRL models implemented in routing, congestion control, chaining of service functions, vehicular communication, and the optimization of optical networks. It also highlights their performance strengths, including topology awareness, cross-topology generalization, high sample efficiency, and high scalability, as well as their weaknesses, such as inference overhead, inconsistent benchmarking practices, low real-time deployability, and sensitivity to noisy or partial state observations. The main findings of this review are: (i) a coherent taxonomy of GNN-based, DRL-based, and hybrid GNN–DRL effective designs; (ii) comparative analysis of algorithms, architecture components, and learning pipelines; (iii) generalized performance trends in major areas of intelligent networking; and (iv) a collection of grounded research directions to be followed in the future, lightweight architecture, transfer learning pipeline, fault tolerant learning, and unified evaluation frameworks. Finally, this review focuses on enabling resilient infrastructure through intelligent, scalable, and autonomous end-to-end networking solutions.