100 REAL TASKS · 6 DOMAINS · HONEST RUN 2026-09-11
WebAgent Failure Atlas
中文真实场景 AI 浏览器 Agent 失败图鉴 — 同一套 harness 真抓每一步 HTTP 与快照 hash,失败即数据。镜像英文基准 WebArena / Mind2Web / VisualWebArena / WebVoyager / OSWorld 的方法学,换成中文真网。
总体通过率 Overall
68.0% 68/100
95% Wilson CI 58.3%–76.3%
最硬墙 / 最稳
政务 29% · 电商 100%
ecom 17/17 · gov 5/17 · χ² p=0.0007
失败 Top2
F6×14 · F3×8
中文表单语义 · 动态渲染未等待
干预实证
1/32 3.1%
5s 退避重试仅救回 search-01 (F3)
🇨🇳 中文
研究亮点
- 100 个中文真实任务:搜索 17 / 电商 17 / 表单 17 / 政务 17 / 地图 16 / 媒体 16
- 总体成功率 68.0%(95% Wilson CI 58.3%–76.3%)
- 政务最硬墙(5/17,29%)vs 电商最稳(17/17,100%)
- 失败 Top2:中文表单语义 F6×14、动态渲染未等待 F3×8
- 重试只救回 3.1%(1/32),证明软墙需结构性修复
- 卡方检验 χ²=21.216,p=0.0007,领域差异极显著
🇺🇸 English
Key Findings
- 100 real Chinese tasks: search 17 / ecom 17 / form 17 / gov 17 / map 16 / media 16
- Overall 68.0% (95% Wilson CI 58.3%–76.3%)
- Gov = hard wall (5/17, 29%) vs Ecom = trivial (17/17, 100%)
- Top failures: F6 Chinese form semantics ×14, F3 dynamic rendering ×8
- Retry rescued only 3.1% (1/32) — soft walls need structural fixes
- Chi-square χ²=21.216, p=0.0007 — domain matters significantly
分域表现 / By Domain
| Domain 领域 | n | pass 通过 | rate 通过率 | 95% CI |
| 🛒 ecom 电商 | 17 | 17 | 100.0% | 81.6%–100.0% |
| 🔍 search 搜索 | 17 | 13 | 76.5% | 52.7%–90.5% |
| 🗺️ map 地图 | 16 | 12 | 75.0% | 50.5%–89.8% |
| 📰 media 媒体 | 16 | 11 | 68.8% | 44.4%–85.8% |
| 📝 form 表单 | 17 | 10 | 58.8% | 36.0%–78.4% |
| 🏛️ gov 政务 | 17 | 5 | 29.4% | 13.3%–53.1% |
失败分类 / Failure Taxonomy (7 classes, 5 observed)
| Class | n | 中文含义 | English Description |
| F6 | 14 | 中文表单语义误解 | Chinese form semantics |
| F3 | 8 | 动态加载未等待 | Dynamic rendering not waited |
| F1 | 4 | 元素定位失败 | Element locator failure |
| F5 | 4 | 验证码/反爬阻断 | Captcha / anti-scrape block |
| F4 | 2 | 登录墙阻断 | Login wall block |
| F2 | 0 | 视觉 grounding 错误 | Visual grounding error |
| F7 | 0 | 多步遗忘/偏离 | Multi-step drift |
🇨🇳 给 Builder 的 5 条修复杠杆
- 先等后断言:动态页固定等待 + 内容长度门限(<20KB 判 JS 壳),能吃掉大半 F3
- 备选选择器:同一语义 ≥3 种定位(文本/role/URL),一个不行换一个
- 登录墙早退:命中关键词即停并标记 F4,不浪费步数
- 反爬退避:403/429 指纹退避 + 降级镜像,保证轨迹不断档
- 中文表单专项:label→input 关联检查,无 label 直接告警 F6
🇺🇸 5 Fix Levers for Builders
- Wait before assert: fixed wait + <20KB JS-shell gate kills most F3
- Fallback selectors: ≥3 per semantic (text/role/URL), rotate on failure
- Login wall early exit: tag F4 immediately, save steps
- Anti-scrape backoff: 403/429 fingerprint + mirror fallback
- Chinese form check: label→input association, alert on missing label
一键复现 / One-click Reproduce
# One-click verify (秒级)
python reproduce.py
# → n=100 pass=68 rate=0.680 CI=[0.5834,0.7633] retry 1/32
# Full rebuild (~2min)
bash verify.sh
# → stats + 9 figs + papers + dashboard
Single source of truth: stats.json → papers / dashboard / landing page all injected from this one file. Numbers are identical across all 4 channels.
单一事实源:stats.json → 论文/仪表盘/落地页均由此注入,四渠道数字完全一致。
诚实边界 / Honest Limits:
中文:规则执行器 = 自动化下界(真实 LLM Agent 只会更高);n=100 CI 仍 ±9%;DOM 漂移 2026-09-11;观察性关联非因果;22 参考文献逐条 HTTP 200。AI 使用声明:代码/分析/文稿有 AI 辅助,科学判断与责任由人类作者承担。
English: Rule-based executor = lower bound (real LLM agents score higher); n=100 CI ±9%; DOM drift 2026-09-11; observational not causal; 22 refs HTTP-verified. AI disclosure: code/analysis/drafting had AI assistance; scientific judgments are human authors'.
方法学致谢 / Methodology Acknowledgments
WebArena (2307.13854) · Mind2Web (2306.06070) · VisualWebArena (2401.13649) · WebVoyager (2401.13919) · OSWorld (2404.07972) — 5 篇标杆奠定方法学,本研究镜像其纪律换成中文真网。
These 5 landmarks established the science of web agent evaluation — we port their discipline to the real Chinese web.
引用 / Cite
@dataset{atlas2026webagent,
title={WebAgent Failure Atlas: 100 Real Chinese Tasks},
year={2026},
url={https://github.com/kevindurant735rocket-creator/webagent-failure-atlas}
}