Attributed graph clustering (AGC) is a prominent research focus in graph data exploration. Although considerable efforts have been made, there still remains two essential issues in recent studies. First, the mainstream AGC approaches mostly derive inspiration from the deep learning models, which may inherit the expensive computation burdens and unnecessary complex structures. Second, they often neglect the potential high-order relationship hidden in graph structural data. In view of this, we develops a lightweight approach termed Multi-scale Multi-order Attributed Graph Clustering (MM-AGC) with nearly linear complexity. Specifically, our approach generates a set of multi-step propagate matrices as multiple views, thereby simultaneously capturing the first-order and high-order topological relationships for subsequent clustering module. Inherited from this foundation, a linear graph filter-based model is purposed with tri-factorization guidance. To be specific, this design seamlessly leverages the linear graph autoencoder and scalable bipartite graph learning into a unified framework. By going beyond the conventional constraints, the multi-view bipartite representations are further extended with multiple diversified anchor sets, which helps to flexibly explore the hierarchical information from multi-scale perspective. Extensive experiments have been conducted to validate the effectiveness and efficiency of our MM-AGC approach against several state-of-the-art competitors.