A new survey examines the relationship between self-evolving AI agents and dynamic graph topology. It frames agent evolution as a process of transforming graphs, where nodes, edges, and subgraphs change over time. The authors propose a taxonomy for existing methods and suggest using dynamic graph learning as infrastructure for self-evolving agents. They also discuss evaluation and governance protocols from a graph-aware perspective. --- Why it matters: This research matters to AI engineers because it provides a new framework for designing and governing self-evolving agents, which are increasingly used in applications such as language models and autonomous systems. Source: https://arxiv.org/abs/2608.18104