arXiv cs.AIThursday · May 28, 2026FREE

SkillGrad: Optimizing Agent Skills Like Gradient Descent

SkillGrad, introduced in arXiv paper 2605.27760, proposes a gradient-descent-inspired framework for optimizing agent skills in LLM agents. Skills are stored as structured files containing procedural knowledge, but they are often unreliable or outdated. SkillGrad treats the skill package as a parameter to optimize: task executions provide trajectory-level loss, automatic diagnoses generate text-based gradients indicating correction directions, and a momentum agent accumulates recurring patterns into persistent memory. An LLM-based patcher then applies layer-aware edits. Evaluated on SpreadsheetBench Verified and WikiTableQuestions, SkillGrad consistently outperforms training-based skill evolution methods. The paper is available on arXiv, published May 28, 2026.