Can AI agents learn and improve their own skills automatically?

Agent skills are usually hand-written or generated once, with no systematic way to improve them. SkillOpt treats skills as trainable external state: a separate optimizer model proposes edits (add/delete/replace lines) to a skill document, accepts only improvements on held-out validation, and uses a textual learning-rate budget to keep training stable. Across 52 model-benchmark combinations, it beats human prompts, one-shot generation, and prior methods like TextGrad by +19–25 points, and skills transfer across model scales and execution environments.