SkillMOO: Multi-Objective Optimization of Agent Skills for Software Engineering

The recently introduced SkillMOO framework, detailed in arXiv paper 2604.09297, conceptualizes software engineering (SE) agent skill bundles as multi-objective search entities. It enhances these bundles via edits suggested by LLMs and employs NSGA-II Pareto selection, focusing on pass rate and inference cost. SkillMOO was tested on all 16 SkillsBench SE tasks, achieving the highest pass rate in 11 out of 12 tasks with non-zero passes, alongside cost reductions of up to 31.7% compared to static bundles and improvements in pass rates by as much as 21 percentage points. An examination of 38 skill edits indicates that pruning and substitution are the most effective strategies, providing valuable insights for skill bundle development. The paper emphasizes that relying solely on static, hand-crafted skills is inadequate, as they may enhance task success while increasing token costs or offering misleading guidance. This framework aims to optimize both pass rates and inference costs, overcoming the shortcomings of existing methods that focus only on pass rates.