SkillOpt: Agent skills as trainable parameters

SkillOpt is a method that treats agent skill files as trainable parameters outside a frozen model, using a training-style optimization loop to improve agent performance without changing model weights. It achieves consistent gains across 52 evaluation cells spanning six benchmarks, seven target models, and three execution modes, outperforming human-written skills and other baseline methods. The optimized skills remain compact, auditable, and transferable across model scales and tasks.

From the source
SkillOpt treats the skill file as a trainable parameter living outside a frozen target model, bringing a training-style optimization loop, consistent gains across 52 evaluation cells, and a compact skill file that stays readable, auditable, and transferable.
microsoft.com

Visit source ↗microsoft.com