Critical nodes in complex networks are of fundamental importance, as they play essential roles in maintaining the network’s functionality, performance, stability, robustness, and dynamic behavior. Identifying critical nodes in complex networks is a significant research topic with broad applications across diverse domains, including biology, sociology, transportation, information science, computer networks, and communication systems. Most existing algorithms for identifying critical nodes rely on predefined mathematical models or heuristic rules to determine target strategies. However, these manually designed, rule-based strategies are constrained by fixed-dimensional representations, which significantly limit their ability to extract meaningful features. To address this issue, this paper proposes a framework named GCRA: Graph Contrastive Reinforcement learning based on Adaptive augmentation. The framework introduces an adaptive graph contrastive learning algorithm that effectively captures both the structural topology and node attributes, thereby representing the overall state of nodes and the network. Subsequently, reinforcement learning is integrated to identify critical nodes by gradually disintegrating the network. The model, trained on randomly generated BA networks, can be directly applied to real-world networks. In experiments, the algorithm is evaluated on synthetic networks of varying scales and real-world networks from diverse domains, and compared against eight benchmark methods. The proposed GCRA achieves performance improvements of 4.91% (unweighted networks) and 14.26% (weighted networks) in critical node identification, demonstrating its effectiveness for complex network analysis.