SkillOpt is a Microsoft Research system that treats natural-language agent skill documents as trainable state and optimizes them through trajectory-driven rollouts, reflection, bounded edits, and validation-gated updates — without touching the frozen LLM weights. It achieves best-or-tied-best performance across all 52 model × benchmark combinations tested.
Text-space optimizer for frozen LLM agents — treats a compact natural-language skill document as the trainable parameter; optimizes it through agent rollouts, reflection steps, and bounded edits without any gradient computation or model fine-tuning
Validation-gated updates for stable skill evolution — each candidate edit is evaluated on a held-out validation set before being committed; prevents skill regression and enables controllable, monotonically improving skill documents across training iterations
Best-or-tied-best across all 52 model × benchmark combos — transfers skills across model families (GPT, Claude, Gemini); skills learned on one agent can be reused by a completely different frozen LLM with no retraining, enabling a composable agent skill ecosystem