Framing Self-Evolving Agents as Dynamic Graph Transformations

This survey proposes a new perspective on LLM agents by modeling agent evolution as dynamic graph transformations. It argues that agent state—including memory, tools, and skills—should be viewed as an evolving topology rather than just a support structure.