SkillGrad: Optimizing Agent Skills Like Gradient Descent

Researchers introduced SkillGrad, a gradient-descent-inspired framework for optimizing LLM agent skills as editable parameters.

It uses execution losses, automatic diagnosis, momentum memory, and LLM-based patching to update skill packages.

On SpreadsheetBench Verified and WikiTableQuestions, SkillGrad beat training-based baselines across two LLM backbones by 6.7 points on average.

The result suggests a more systematic way to improve agent skills and outperform prior skill-evolution methods on benchmarks.