SkillOpt

Most agent skills are hand-written once and then left alone. SkillOpt from Microsoft Research treats a skill markdown file as the trainable state of a frozen model: scored rollouts become bounded edits, validation gates accept only real gains, and the deploy artifact is a compact best_skill.md with zero extra model calls at runtime. Across many model and harness cells it lifted no-skill baselines hard, including big jumps inside Codex and Claude Code. The useful shift is skill writing as controlled optimization, not another prompt pack.