The classic and widely used Graph Attention Network (GAT) is designed to generate embedding representations for graph-structured data via the attention-based neighborhood aggregation (a.k.a. message passing) mechanism. The recently proposed Kolmogorov-Arnold Network (KAN) has demonstrated superior learning capabilities compared to Multi-Layer Perceptron (MLP) across numerous tasks. This study explores the integration of KAN into the vanilla GAT to enhance feature extraction capabilities and proposes five different KAGAT variants. Node classification experiments conducted on real-world datasets indicate that the proposed KAGATs outperform classic GNNs such as GCN, GAT, GATv2, and GIN, as well as other existing KAN-based GNNs. Our code is available at https://github.com/Xkzysheng/Kolmogorov-Arnold-Graph-Attention-Network .