SkillOpt

Microsoft SkillOpt optimizes agent skills via text-space training

Microsoft researchers published SkillOpt, a systematic text-space optimizer that trains agent skills as external frozen-model state using controlled edit loops, achieving +23.5 point accuracy gains on GPT-5.5 across six benchmarks with zero inference overhead. Commerce teams deploying agentic workflows gain a reproducible skill-tuning method that transfers across model scales and execution environments without retraining.