Webwright：微软开源的浏览器 Agent，Code-as-Action 在 Mind2Web 上拿到 86.7%

📌 开源仓库：microsoft/Webwright GitHub：https://github.com/microsoft/Webwright License：MIT | Stars：5,996 作者：Yadong Lu, Lingrui Xu, Chao Huang, Ahmed Awadallah（微软）

---

大多数浏览器 Agent 做的事是：截图 → 识别位置 → 预测坐标 → 点击。

Webwright 的做法不同：把浏览器当作代码执行的对象，而不是感知的对象。 模型写 Playwright Python 脚本，脚本操作浏览器，结果写回本地工作区——不用每步截图，不用每步打坐标。

---

一、为什么 Code-as-Action 比坐标点击强

坐标点击（screenshot + xy-coordinate）的问题：

每一步都要截图、分析、生成坐标、点击、再截图——这是一条线性链，中间任何一环出错（识别偏移、元素遮挡、DOM 变化）就从头来过，而且每步都消耗视觉 token。

Webwright 的做法：

模型写一段 Playwright 代码，直接描述操作序列——page.click("#submit-button")、page.fill("input[type=email]", value) 这类语义操作，而不是”点击坐标 (342, 518)“。代码可以重跑、可以调试、可以复用。

更关键的是：本地工作区（代码 + 日志）是状态，不是浏览器会话。这意味着：

• 脚本可以跨任务复用
• Agent 重启后工作区不丢失
• 多步操作可以在一次代码执行里完成

---

二、基准测试结果

Online-Mind2Web（300 个真实网页任务）：

模型	准确率
GPT-5.4	86.7%
Claude Opus 4.7	84.7%

Odysseys（200 个长时域任务）：

模型	准确率
GPT-5.4	60.1%

这个 60.1% 比此前 SOTA 高 +15.6 个百分点，且 Odysseys 专门设计来测”做完一件需要多步骤的事”——对真实 Agent 场景更有参考价值。

WebArena + Skill Factory：

• 基础准确率：55%
• 启用 Skill Factory 后：70%（+15 pp）

Skill Factory 是后面会详细说的机制，意义在于：已经解过的任务不用再让模型重新想——复用代码，零 token 消耗。

---

三、代码规模：极简

这不是一个塞满功能的大框架：

组件	行数
核心 agent 循环	~450
Playwright 环境	~570
CLI 接口	~150

依赖只有四个：httpx、pydantic、playwright、typer。

这个规模的好处是：可以完整读完，可以 fork 改造，可以嵌进别的系统，不会被框架细节淹没。

---

四、Skill Factory：解过的任务不再重算

Skill Factory 是 Webwright 里设计最独特的部分。

工作流：

1. Agent 完成一个任务，生成了 Playwright 脚本
2. 该脚本被参数化、封装成独立 Skill
3. 下次遇到类似任务，直接运行 Skill——约 40 秒，零 token

在 WebArena 上，这个机制把准确率从 55% 提升到 70%。

背后的逻辑是：浏览器上很多任务是高度重复的——登录、搜索、填表、提交——第一次让 Agent 完整思考，之后都走确定性脚本。这是用代码形态做的”经验记忆”。

---

五、架构和支持后端

webwright/
├── agents/default.py       # 核心 agent 循环（~450 行）
├── environments/           # Playwright 工作区
├── models/                 # OpenAI / Anthropic / OpenRouter 后端
└── skills/webwright/       # Claude Code / Codex / OpenClaw / Hermes 插件清单

支持的模型后端：

• OpenAI：GPT-5.4 及以上
• Anthropic：Claude Opus 4.7、4.6
• OpenRouter：任意兼容模型

插件集成：已提供 Claude Code、Codex、OpenClaw、Hermes 的 skill 清单，可以直接作为这些平台的插件调用，不需要额外 API 成本。

---

快速开始

pip install -e .
playwright install chromium

python -m webwright.run.cli \
    -c base.yaml -c model_openai.yaml \
    -t "在 Amazon 上搜索机械键盘并找到价格最低的" \
    --start-url https://amazon.com \
    --task-id demo_01 \
    -o outputs/default

---

拆解结论

Webwright 在几个维度上都值得关注：

数字上：Mind2Web 86.7%、Odysseys +15.6pp，是当前公开基准里排得上的成绩。

工程上：~450 行核心循环，极简依赖——这是可以读懂、可以改造的规模。

思路上：Code-as-Action vs 坐标点击，Skill Factory 的”已解任务变代码”——这两个设计思路在其他 Agent 框架里不多见。

对比 Stagehand、browser-use 这些框架，Webwright 的差异在于：把本地工作区当状态，而不是浏览器会话当状态，这让任务经验可积累、可复用。

---

开源代码与模型仅供学习、勿直接用于工作。

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

📌 Repository: microsoft/Webwright GitHub: https://github.com/microsoft/Webwright License: MIT | Stars: 5,996 Authors: Yadong Lu, Lingrui Xu, Chao Huang, Ahmed Awadallah (Microsoft)

---

Most browser agents work like this: screenshot → detect position → predict coordinates → click.

Webwright does something different: treat the browser as an object of code execution, not an object of perception. The model writes Playwright Python scripts; scripts operate the browser; results are written back to the local workspace — no per-step screenshots, no per-step coordinates.

---

I. Why Code-as-Action Beats Coordinate Clicking

The problem with screenshot + xy-coordinate:

Every step requires a screenshot, analysis, coordinate generation, a click, and another screenshot — a linear chain where any link failure (detection offset, element occlusion, DOM change) means starting over, and every step burns visual tokens.

Webwright’s approach:

The model writes a Playwright script that directly describes the operation sequence — page.click("#submit-button"), page.fill("input[type=email]", value) — semantic operations, not “click coordinate (342, 518).” Code can be re-run, debugged, and reused.

The key insight: the local workspace (code + logs) is the state, not the browser session. This means:

• Scripts are reusable across tasks
• The workspace survives agent restarts
• Multi-step operations complete in a single code execution

---

II. Benchmark Results

Online-Mind2Web (300 real web tasks):

Model	Accuracy
GPT-5.4	86.7%
Claude Opus 4.7	84.7%

Odysseys (200 long-horizon tasks):

Model	Accuracy
GPT-5.4	60.1%

That 60.1% is +15.6 points over prior SOTA. Odysseys is specifically designed to test “completing something that requires many steps” — more relevant to real-world agent scenarios than click-accuracy benchmarks.

WebArena + Skill Factory:

• Base accuracy: 55%
• With Skill Factory: 70% (+15 pp)

---

III. Codebase Size: Genuinely Minimal

Component	Lines
Core agent loop	~450
Playwright environment	~570
CLI interface	~150

Four dependencies: httpx, pydantic, playwright, typer.

At this scale, you can read it end to end, fork and modify it, embed it into other systems — without getting buried in framework internals.

---

IV. Skill Factory: Never Solve the Same Task Twice

Skill Factory is Webwright’s most distinctive design.

Workflow:

1. Agent completes a task, produces a Playwright script
2. That script is parameterized and packaged as a standalone Skill
3. Next time a similar task appears, run the Skill — ~40 seconds, zero tokens

On WebArena, this mechanism lifted accuracy from 55% to 70%.

The logic: many browser tasks are highly repetitive — login, search, fill, submit. The agent thinks it through once, then all future instances run deterministic scripts. It’s “experience memory” in code form.

---

V. Architecture and Supported Backends

webwright/
├── agents/default.py       # core agent loop (~450 lines)
├── environments/           # Playwright workspace
├── models/                 # OpenAI / Anthropic / OpenRouter backends
└── skills/webwright/       # plugin manifests for Claude Code / Codex / OpenClaw / Hermes

Supported model backends:

• OpenAI: GPT-5.4 and above
• Anthropic: Claude Opus 4.7, 4.6
• OpenRouter: any compatible model

Plugin integrations: Skill manifests already provided for Claude Code, Codex, OpenClaw, and Hermes — call Webwright directly as a plugin without extra API costs.

---

Quick Start

pip install -e .
playwright install chromium

python -m webwright.run.cli \
    -c base.yaml -c model_openai.yaml \
    -t "Search for mechanical keyboards on Amazon and find the lowest price" \
    --start-url https://amazon.com \
    --task-id demo_01 \
    -o outputs/default

---

Teardown Summary

Webwright stands out on several dimensions:

Numbers: Mind2Web 86.7%, Odysseys +15.6pp — leading scores on current public benchmarks.

Engineering: ~450-line core loop, minimal dependencies — a codebase you can actually read and modify.

Design: Code-as-Action vs. coordinate clicking, and Skill Factory’s “solved tasks become code” — two ideas you won’t find widely implemented in other browser agent frameworks.

Compared to Stagehand, browser-use, and similar frameworks, Webwright’s key differentiator is: local workspace as state, not browser session as state — making task experience accumulate and remain reusable.

---

Open-source code and models are for learning purposes only — do not use directly in production work.

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.