GraphSkillEvo: Evolutionary Optimization of Graph-Structured Agent Skills

This paper addresses the limitation of unstructured natural-language instructions for LLM agent skills, which lack explicit workflow guidance and create an unmanageable search space. The authors propose representing skills as graph-structured natural-language artifacts, where each node is an execution step and directed edges encode context-dependent transitions. Building on this representation, they introduce GraphSkillEvo, a population-based evolutionary optimization framework featuring mutation and crossover operators for exploring the structured skill space. Experiments across five benchmarks show consistent performance improvements over the SkillOpt baseline.

GraphSkillEvo consistently outperforms the strong skill optimization baseline SkillOpt, improving average accuracy by 4.01% on GPT-5.4-nano and 1.76% on GPT-5.4.