SkillOpt
Microsoft Research treats a compact natural-language skill document as the trainable state of a frozen agent, then learns that document through rollouts, reflection, and bounded edits gated by held-out validation. The argument is direct: most engineers handwrite agent skill docs and hope they generalize, when the doc itself should be optimized like a parameter. SkillOpt reframes the SKILL.md file as an external parameter of a model whose weights never change.
Ask this paper
The skill doc as a trainable parameter: An optimizer model proposes validation-gated edits to the skill file, adding, deleting, or replacing instructions. A textual learning rate controls how aggressively each round rewrites the document, with batch and momentum reported in text space rather than gradient space.
Validation gates instead of hope: Every edit must pass a held-out check before it is kept. This turns skill authoring into a measurable optimization loop with a real objective, rather than prompt tweaking guided by intuition.
52 out of 52 wins: SkillOpt beats Trace2Skill, TextGrad, GEPA, EvoSkill, human-written skills, and one-shot skills across 6 benchmarks and 7 target models. It adds roughly +23.5 points on GPT-5.5 in direct chat, +24.8 in the Codex loop, and +19.1 in Claude Code from the no-skill baseline.
Why it matters: If the skill document is the thing you optimize, the bottleneck shifts from base-model capability to how well you can train the natural-language state around a frozen agent. That is a cheap, model-agnostic lever most teams are leaving on the table.