Current stereo methods grapple with the challenges of extensive high computational complexity, primarily due to the formulation of dense cost/attention volume. To tackle this problem, we propose an iterative aggregation on a dynamic graph structure, where the nodes correspond to the attribute of pixels, and the edges capture their information propagation flow. This adaptive framework grants each node the flexibility to customize its own aggregation scope aligned with its individual matching conditions, thereby encouraging the aggregation workload toward the critical regions. We further bridge the graph-based aggregation and disparity prediction via correspondent embeddings, which progressively capture the matching cues on graphs and are ultimately mapped into disparity maps through a sequential decoder. Experimental results on standard benchmarks demonstrate that our method achieves real-time efficiency compared to state-of-the-art techniques while maintaining competitive accuracy.