SkillOpt 怎么训练 Agent Skill：从 SKILL.md 到 best_skill.md 的工程流程

Microsoft SkillOpt gives agent teams a concrete way to train skill documents without touching model weights. The target model stays frozen; a separate optimizer reads scored trajectories and proposes bounded edits to the skill file. A candidate skill is accepted only when it improves held-out validation. That turns SKILL.md from hand-written advice into an artifact with a training loop, score history, and deployment boundary.

The practical reason this matters is simple: agents already accumulate instructions. Teams add rules after failures, append caveats after edge cases, and copy working prompts between projects. Without a gate, those changes can become prompt debt. SkillOpt keeps the useful part of that habit—the document can learn from failures—while adding the missing discipline: edit budgets, validation splits, and rejected candidates.

What to evaluate before using it

• Do you have repeatable tasks and real scores, not only anecdotes?
• Can you hold out validation examples that the optimizer does not train on?
• Will someone review proposed skill changes before they affect production workflows?
• Is the target workflow stable enough for a trained skill to remain useful?

The new SkillOpt-Sleep preview is also notable. It can harvest supported coding-agent sessions, mine recurring tasks, replay them offline, and stage memory or skill updates for review. The documentation is explicit about privacy: real backends may send excerpts and derived tasks to the chosen provider, so sensitive teams need inspection and redaction before replay.

Used carefully, SkillOpt is not a magic self-improving agent. It is a way to make agent learning boring in the right sense: measured, gated, reviewable, and reversible.