SkillOpt: Agent skills as trainable parameters

AI agents often struggle due to manual skill modifications that lack guaranteed improvements. SkillOpt transforms skill editing into a training process, making agent behavior more reliable without altering model weights. It excels across six benchmarks, achieving top performance in all 52 evaluation cells. By treating the skill file as a trainable parameter, SkillOpt enhances performance through controlled optimizations rather than random edits. This ensures skills remain compact and auditable, preventing uncontrolled prompt drift.