SkillOpt

SkillOpt is a research framework from Microsoft that treats an agent's skill document as trainable state. Instead of fine-tuning weights, it runs a training loop — rollout, reflect, aggregate, select, update, evaluate — over the markdown a frozen model reads, and borrows the machinery that makes weight-space optimisation reproducible: epochs, mini-batches, a textual learning-rate budget, a rejected-edit buffer, and epoch-wise slow/meta updates.

How the loop works

A separate optimizer model converts scored rollouts into bounded add / delete / replace edits on one skill document. In the default path an edit is only accepted when it strictly improves a held-out validation score, which is what stops the loop from drifting into confident nonsense — the failure mode of "let an LLM rewrite its own prompt" approaches.

The deployed artifact is a compact best_skill.md, typically 300–2,000 tokens, that runs against the unchanged target model and adds no inference-time model calls.

Reported results

Across six benchmarks, seven target models and three execution harnesses — direct chat, Codex CLI and Claude Code CLI — the authors report SkillOpt best or tied-best on all 52 evaluated cells. On GPT-5.5 they measure average accuracy lifts over no-skill baselines of +23.5 points in direct chat, +24.8 in the Codex agentic loop and +19.1 in Claude Code, with optimised artifacts transferring across model scales, between harnesses, and to nearby benchmarks. The method, ablations and per-cell numbers are in the arXiv paper (2605.23904).

Using it

pip install skillopt. It ships multi-backend support for OpenAI, Azure, Claude, Qwen, MiniMax and any OpenAI-compatible endpoint, plus exec backends for Codex, Claude Code, Cursor and Copilot; six built-in benchmarks; and an optional Gradio monitoring dashboard. Version 0.2.0 added SkillOpt-Sleep, a nightly offline self-evolution engine that harvests past sessions, replays recurring tasks and consolidates skills behind the same held-out gate, exposed as a skillopt-sleep CLI, together with integration shells for Claude Code, Codex, Copilot and Devin.

MIT licensed. Python 3.10+. Best suited to teams who already measure their agents and want a disciplined way to improve the instructions rather than guessing at prompt rewrites.