SkillOpt: First Systematic Text-Space Optimizer Treats Agent Skills Like Trainable Parameters

arXiv·high signal

SkillOpt proposes treating agent skills as external state of a frozen LLM, optimized with the same discipline as weight-space training. Rather than hand-crafting, one-shot generating, or loosely self-revising skills, SkillOpt applies controllable text-space optimization that reliably improves skill performance over its starting point under feedback. This is the first framework to bring reproducible optimization methodology to the skill-text layer of agent systems.