SkillOpt-Lite: Better and Faster Agent Self-Evolution via One Line of Vibe

SkillOpt-Lite frames agent skill optimization as zeroth-order optimization and reduces the pipeline to three components: file-system-based trajectory exploration, consensus attribute mining, and independent validation gating. The paper reports LiveMath gains of 8.8 points on GPT-5.5 and 25.4 points on GPT-5.4-nano versus full SkillOpt. Its broader HarnessOpt extension reportedly reaches 0.7758 accuracy for GPT-5.4-nano on SpreadsheetBench, exceeding GPT-5.5 at 0.7620 under standard pipelines. The authors also describe integration with coding agents such as VSCode Copilot.