Attributed graphs are essential for modeling complex relational data across numerous domains. Existing clustering methods typically address embedding and clustering separately, while joint approaches often rely on fixed, non-learnable update steps. Both strategies limit the adaptability and integration of these methods into broader learning pipelines. In this work, we introduce E-PAGEC, Enhanced Power Attributed Graph Embedding and Clustering, a differentiable variant of PAGEC that jointly refines node embeddings and cluster assignments through gradient-based optimization. In contrast to prior non-adaptive designs, E-PAGEC incorporates topology, similarity of attributes, and iterative updates driven by gradients, enabling the model to dynamically adjust both embedding space and cluster boundaries during training. Experiments on real-world datasets, against state-of-the-art graph clustering baselines, demonstrate that E-PAGEC achieves superior clustering performance.