SkillOpt：微软联合上交复旦同济，让 AI Agent 的 Skill 像神经网络一样被训练

Skill 能运行，不代表 Skill 写得好。

这是 SkillOpt 诞生的出发点。你花时间为 AI Agent 写了一份 Skill 文档，但你怎么知道这份 Skill 已经是最优的？或者，它还有多大提升空间？

2026 年 5 月，微软研究院联合上海交通大学、复旦大学、同济大学发布了 SkillOpt（arXiv:2605.23904），给出了一个系统化的答案：把 Skill 文档当作神经网络的参数来训练。

GitHub 已达 5278 star、532 fork，PyPI 上线即可 pip install。

问题：Skill 的三条死路

目前主流的 Agent Skill 写法有三种，SkillOpt 的论文明确指出它们都有根本性缺陷：

方式	缺陷
手写 Skill	凭经验，缺乏系统反馈，无法可靠提升
一次性 LLM 生成	没有反馈循环，生成即定格
松散自修订	缺乏控制，方向随机，可能越改越差

SkillOpt 的核心洞察是：Skill 文档应该是可训练的状态，而不是静态产物。

核心思路：文本空间的深度学习类比

SkillOpt 把深度学习的训练范式完整移植到了文本空间：

深度学习              SkillOpt（文本空间）
─────────────────────────────────────────────
参数（Parameter）  →  Skill 文档
梯度（Gradient）   →  轨迹推导出的编辑方向
学习率（LR）       →  每步允许的编辑预算
验证集（Valid.）   →  Hold-out 验证门控
批次（Batch）      →  一批执行轨迹（Minibatch）
Epoch              →  一轮完整优化周期

最关键的是：模型权重完全冻结，优化的只有 Skill 文档本身。你不需要微调模型，不需要 GPU，只需要 API 调用权限。

训练循环：六步迭代

每一轮 SkillOpt 的工作如下：

1. Rollout  → 用当前 Skill 执行任务，记录轨迹和分数
2. Reflect  → 分析一批轨迹，提炼成功/失败模式
3. Aggregate→ 将反思汇总为有界的文本编辑（增/删/改）
4. Select   → 候选编辑必须通过验证集，严格改善才接受
5. Update   → 接受的编辑写入 Skill，被拒编辑进入负反馈缓冲区
6. Evaluate → 在测试集上评估，输出 best_skill.md

三个稳定性机制保证不会”越改越差”：

• 验证门控：候选 Skill 必须在 hold-out 集上严格涨分才被接受
• 有界编辑：每步只允许 4-8 个修改（类似学习率限制）
• 慢/元更新：Epoch 级别的跨批次方向稳定机制

实验结果：52 局全胜

这是 SkillOpt 最让人印象深刻的地方。

跨 6 个基准测试、7 个目标模型、3 种执行环境（直接对话、Codex CLI、Claude Code CLI），共 52 个评测单元，SkillOpt 全部获胜或并列最优。

对比对象包括：手写 Skill、一次性 LLM 生成、Trace2Skill、TextGrad、GEPA、EvoSkill。

以 GPT-5.5 为例的提升幅度：

执行环境	平均准确率提升
直接对话	+23.5 分
Codex 代理循环	+24.8 分
Claude Code	+19.1 分

部分 benchmark 的具体数字更震撼：

• SpreadsheetBench：+38.9 分
• OfficeQA：+39.0 分

更重要的是，迁移性表现优秀：用 GPT-5.4 优化出的 Skill，迁移到小模型后仍保留约 82% 的增益；从 Codex 迁移到 Claude Code 环境，无需重新优化。

这意味着：一份优化好的 best_skill.md，可以跨模型、跨平台复用。

普通开发者怎么用

安装（5 分钟）

pip install skillopt

# 如果要用 Gradio 监控界面：
pip install skillopt[webui]

# Claude 后端：
pip install skillopt[claude]

配置 API

cp .env.example .env
# 编辑 .env，填入 API 密钥

# OpenAI 兼容接口（OpenAI / DeepSeek / 任意 v1 接口）：
export AZURE_OPENAI_ENDPOINT="https://api.openai.com/v1"
export AZURE_OPENAI_API_KEY="sk-..."
export AZURE_OPENAI_AUTH_MODE="openai_compatible"

# Claude：
export ANTHROPIC_API_KEY="sk-ant-..."

# Qwen（本地 vLLM）：
export QWEN_CHAT_BASE_URL="http://localhost:8000/v1"
export QWEN_CHAT_MODEL="Qwen/Qwen3.5-4B"

运行优化（核心三步）

Step 1：准备你的 Skill（或者从空白开始）

# my_skill.md（种子 Skill，可以很简略）
你是一个专注于数学推理的 AI，每一步都要写出完整推理过程。

Step 2：开始训练

python scripts/train.py \
  --config configs/searchqa/default.yaml \
  --optimizer_model gpt-4o \
  --target_model gpt-4o-mini

你可以用强模型（gpt-4o）做优化器，用弱模型（gpt-4o-mini）做目标 Agent，最大化性价比。

Step 3：取出成果

训练结束后，ckpt/ 目录下会生成 best_skill.md。这个文件就是优化后的 Skill，300-2000 token，直接复制到你的 Agent 的 System Prompt 里即可。

零额外推理开销——部署时不需要额外调用任何模型，就是一段 Markdown 文本。

监控训练过程

# 启动 WebUI 看实时优化曲线
pip install skillopt[webui]
opensquilla gateway run  # 或按项目文档启动

我的判断

这是一篇方法论严谨、工程落地扎实的研究，值得认真对待。

SkillOpt 解决的不是”能不能用 AI 写 Skill”的问题（那个已经解决了），而是”怎么系统地、可重复地把 Skill 写好”的问题。这个问题在 AI Agent 进入工程化落地阶段之后，变得越来越重要。

对普通开发者最实际的价值：

1. 不需要懂深度学习：整套流程是 API 调用，没有 GPU，没有梯度
2. 结果是可读的 Markdown：你知道 Skill 改了什么，可以审阅、可以 Git 管理
3. 跨模型复用：在贵模型上优化一次，迁移到便宜模型继续用
4. 已有生态接入：gbrain、darwin-skill 等工具已在 2026 年 6 月集成

需要注意：

• 优化需要消耗 API token（优化器 + 目标模型都在跑），有成本
• 效果高度依赖你的任务定义和验证集质量——垃圾进、垃圾出
• 目前 v0.1.0，生产稳定性仍需观察

---

如果你在用 Claude Code、Codex、或任何支持 Skill 文件的 AI Agent，SkillOpt 值得认真试一试。

GitHub： microsoft/SkillOpt
arXiv 原文： 2605.23904 — SkillOpt: Executive Strategy for Self-Evolving Agent Skills
项目主页： https://aka.ms/skillopt
PyPI： pip install skillopt

SkillOpt: Microsoft + Top Chinese Universities Train Agent Skills Like Neural Networks

A skill that runs is not the same as a skill that’s well-written.

That’s the starting point for SkillOpt. You spend time writing a skill document for your AI agent — but how do you know it’s optimal? How much headroom is there for improvement?

In May 2026, Microsoft Research, in collaboration with Shanghai Jiao Tong University, Fudan University, and Tongji University, released SkillOpt (arXiv:2605.23904): a systematic answer. Treat the skill document as a trainable parameter — just like neural network weights.

The GitHub repo already has 5,278 stars and 532 forks, and it’s a pip install away.

The Problem: Three Dead Ends for Agent Skills

The three dominant approaches to writing agent skills all have fundamental flaws, as the SkillOpt paper explicitly notes:

Approach	Problem
Hand-crafted skills	Experience-driven, no systematic feedback, unreliable improvement
One-shot LLM generation	No feedback loop — generated and frozen
Loose self-revision	Uncontrolled direction, can degrade rather than improve

SkillOpt’s core insight: skill documents should be trainable state, not static artifacts.

Core Concept: Deep Learning Analogized to Text Space

SkillOpt ports the full deep learning training paradigm into text space:

Deep Learning         SkillOpt (Text Space)
──────────────────────────────────────────────
Parameter         →   Skill document
Gradient          →   Trajectory-derived edit direction
Learning rate     →   Edit budget per step
Validation set    →   Hold-out selection gate
Batch             →   A batch of execution trajectories
Epoch             →   One full optimization cycle

Crucially: model weights remain completely frozen — only the skill document is optimized. No fine-tuning. No GPU. Just API calls.

The Training Loop: Six Steps

Each SkillOpt round works as follows:

1. Rollout    → Execute tasks with current skill; record trajectories and scores
2. Reflect    → Analyze a batch of trajectories; identify success/failure patterns
3. Aggregate  → Distill into bounded text edits (add / delete / replace)
4. Select     → Candidate edit must strictly improve validation score to be accepted
5. Update     → Accepted edits written to skill; rejected edits enter negative feedback buffer
6. Evaluate   → Score on test set; export best_skill.md

Three stability mechanisms prevent “editing to worse”:

• Validation gating: strict improvement on hold-out set required
• Bounded edits: only 4-8 modifications per step (analogous to learning rate)
• Slow/meta updates: cross-batch directional stability at epoch boundaries

Results: 52-for-52

This is SkillOpt’s most striking result.

Across 6 benchmarks, 7 target models, and 3 execution harnesses (direct chat, Codex CLI, Claude Code CLI) — 52 evaluated (model, benchmark, harness) cells — SkillOpt wins or ties best on every single one.

Competitors included: hand-crafted skills, one-shot LLM generation, Trace2Skill, TextGrad, GEPA, and EvoSkill.

Accuracy gains for GPT-5.5:

Harness	Average Accuracy Gain
Direct chat	+23.5 points
Codex agentic loop	+24.8 points
Claude Code	+19.1 points

Specific benchmark highlights:

• SpreadsheetBench: +38.9 points
• OfficeQA: +39.0 points

Transferability is also strong: a skill optimized on GPT-5.4 retains ~82% of its gains when transferred to a smaller model. Codex-optimized skills transfer to Claude Code without re-optimization.

One optimized best_skill.md can serve across models and platforms.

How to Use It: A Practical Developer Guide

Install (5 minutes)

pip install skillopt

# With Gradio monitoring dashboard:
pip install skillopt[webui]

# Claude backend:
pip install skillopt[claude]

Configure API Access

cp .env.example .env

# OpenAI-compatible (OpenAI / DeepSeek / any v1 endpoint):
export AZURE_OPENAI_ENDPOINT="https://api.openai.com/v1"
export AZURE_OPENAI_API_KEY="sk-..."
export AZURE_OPENAI_AUTH_MODE="openai_compatible"

# Anthropic Claude:
export ANTHROPIC_API_KEY="sk-ant-..."

# Qwen (local vLLM):
export QWEN_CHAT_BASE_URL="http://localhost:8000/v1"
export QWEN_CHAT_MODEL="Qwen/Qwen3.5-4B"

Run Optimization (Three Steps)

Step 1: Prepare your seed skill (or start empty)

# my_skill.md — a minimal seed is fine
You are an AI focused on mathematical reasoning.
Show your full reasoning process at every step.

Step 2: Launch training

python scripts/train.py \
  --config configs/searchqa/default.yaml \
  --optimizer_model gpt-4o \
  --target_model gpt-4o-mini

You can use a strong model (gpt-4o) as the optimizer and a weaker model (gpt-4o-mini) as the target agent to maximize cost efficiency.

Step 3: Collect the result

After training, ckpt/ contains best_skill.md — a 300–2,000 token optimized skill document. Drop it directly into your agent’s system prompt.

Zero inference-time overhead — deployment is just a Markdown text file, no additional model calls required.

My Assessment

Methodologically rigorous, engineered for real deployment — worth serious attention.

SkillOpt isn’t solving “can AI write a skill?” (that was already solved). It’s solving “how do you systematically and reproducibly make a skill good?” — a question that becomes critical as AI agents move from experiments into production.

Practical value for developers:

1. No ML expertise needed: the entire loop is API calls — no GPU, no gradients
2. Readable output: the skill is Markdown — reviewable, versionable, auditable
3. Cross-model reuse: optimize once on an expensive model, transfer to a cheaper one
4. Active ecosystem: gbrain, darwin-skill, and gbrain-evals already integrated (June 2026)

Things to watch:

• Optimization costs API tokens from both optimizer and target model
• Quality depends heavily on your task definition and validation set — garbage in, garbage out
• v0.1.0 is early; production stability is still accumulating

---

If you’re using Claude Code, Codex, or any AI agent that supports skill files, SkillOpt is worth a serious half-day trial.

GitHub: microsoft/SkillOpt
arXiv paper: 2605.23904 — SkillOpt: Executive Strategy for Self-Evolving Agent Skills
Project page: https://aka.ms/skillopt
Install: pip install skillopt