Research proposes new way to model self-evolving AI agents

Researchers from the University of New South Wales have published a survey proposing a novel formalism for self-evolving AI agents: treating agent evolution as dynamic graph transformation. The paper, titled 'Self-Evolving Agents as Dynamic Graph Transformation' and published on arXiv (2608.18104), argues that current approaches either focus on graph structures as static supports or agent-level mechanisms without considering how agent states and graph topologies co-evolve.

The team introduces a modeling framework where agent states—including memories, tools, skills, workflows, and inter-agent relations—are represented as typed nodes, edges, and subgraphs that update through schema-constrained rewrites. This formulation organizes existing methods into four taxonomies: node/feature evolution, edge/topology evolution, subgraph activation, and cross-component co-evolution. The researchers then map nine dynamic graph learning subfields to specific agent capabilities, outlining potential failure modes and proposing graph-aware evaluation protocols.