SkillMOO: Automated Multi-Objective Optimization of Agent Skill Bundles via NSGA-II

arXiv·high signal

Gong, Gu, and Fei present SkillMOO, a framework that automatically evolves agent skill bundles using LLM-proposed edits and NSGA-II survivor selection to balance success rate, cost, and runtime — replacing expensive manual tuning. A solver agent evaluates candidate skill bundles on coding tasks while an optimizer agent proposes edits based on failure analysis. Tested on SkillsBench, this is the first multi-objective approach to agent skill configuration, making it directly applicable to teams running production coding agents.