【Agent工坊】浏览器自动化三件套完全指南

需提取网页正文？ ├── 普通网页 → 优先用web_extract（更快省token） │   └── 失败 → 用browser_console提取 ├── GitHub页面 → 直接用browser_console（web_extract必失败） ├── The Verge/媒体站 → 用browser_console └── Cloudflare拦截 → Web Archive + browser_console  需分析页面结构？ └── 用browser_snapshot（仅调试时使用） 

browser_navigate(url="https://github.com/NousResearch/hermes-agent/releases") 

browser_console(expression="document.querySelector('.markdown-body')?.innerText || document.querySelector('article')?.innerText || document.body.innerText.substring(0, 5000)") 

Hermes Agent v2026.4.16 ========================  ## What's New 🚀  - **Browser Console Tool**: New `browser_console` tool for executing   JavaScript in loaded pages. Supports Promises for async extraction. - **Patch v2**: Fuzzy matching now uses 9 strategies instead of 3.   Minor whitespace differences no longer break replacements. - **Terminal PTY**: Fixed SIGTERM handling in pseudo-terminal mode.   Interactive CLIs (Codex, Claude Code) are now stable.  ## Bug Fixes 🐛  - `read_file` offset parameter: Fixed off-by-one error when offset > 1 - `process.wait`: No longer hangs when process exits during poll cycle - `web_search`: Added bocha as native backend option for Chinese queries  ## Breaking Changes ⚠️  - `browser_snapshot` now returns accessibility tree by default instead   of raw HTML. Use `browser_console` for content extraction. 

# 步骤1：打开AI专栏首页 browser_navigate(url="https://www.theverge.com/ai-artificial-intelligence")  # 步骤2：首页提取最新文章列表 browser_console(expression="Array.from(document.querySelectorAll('[data-analytics-link] h3, .duet--article--headline a')).slice(0, 5).map(el => ({title: el.innerText?.trim(), href: el.href || el.closest('a')?.href}))")  # 返回示例： # [ #   {"title": "OpenAI quietly launched a new moderation tool", "href": "https://www.theverge.com/2026/5/14/..."}, #   {"title": "Anthropic's Claude can now browse the web", "href": "https://www.theverge.com/2026/5/14/..."}, #   ... # ]  # 步骤3：进入目标文章 browser_navigate(url="https://www.theverge.com/2026/5/14/openai-quietly-launched-moderation-tool")  # 步骤4：console提取正文 browser_console(expression="document.querySelector('.duet--article--article-body-component')?.innerText || document.querySelector('[class*=article-body]')?.innerText || document.body.innerText.substring(0, 10000)") 

OpenAI quietly launched a new moderation tool that could reshape how AI companies handle content safety. The tool, called "Guardian," was rolled out to enterprise customers last week without a formal announcement.  According to an internal document obtained by The Verge, Guardian uses a multi-layered approach...  "We believe safety should be invisible but effective," said OpenAI's head of trust and safety.  Industry analysts see this as a direct response to... 

# ❌ 直接访问触发Cloudflare browser_navigate(url="https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified")  # ✅ 使用Web Archive快照 browser_navigate(url="https://web.archive.org/web/20260514/https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified") 

# 通过HN Algolia获取文章objectID curl -s "https://hn.algolia.com/api/v1/search?query=关键词&tags=story&hitsPerPage=5" # 用objectID拉取完整评论 curl -s "https://hn.algolia.com/api/v1/items/44569445" 

# === 内容监控流水线（Hermes Agent Cron任务） ===  # 步骤1：打开HN首页 browser_navigate(url="https://news.ycombinator.com/")  # 步骤2：提取热榜前15条 browser_console(expression="Array.from(document.querySelectorAll('.titleline > a')).slice(0, 15).map(a => ({title: a.innerText, url: a.href}))")  # 返回示例： # [ #   {"title": "Show HN: I built an AI agent that manages my calendar", "url": "https://..."}, #   {"title": "Claude Code can now deploy to production", "url": "https://..."}, #   {"title": "Hermes Agent v2026.5.14 released with browser tools", "url": "https://..."}, #   ... # ]  # 步骤3：Hermes对话中自然语言筛选 # "提取含AI/Agent/Claude/Hermes/LLM的条目，过滤广告"  # 步骤4：对每条目提取正文 browser_navigate(url="具体文章URL") browser_console(expression="document.body.innerText.substring(0, 5000)") # 记录至草稿文件  # 步骤5：生成汇总报告 # 写入content/draft-YYYYMMDD-HHMM-ai-fengxiang-xxx.md 

browser_console(expression="Array.from(document.querySelectorAll('.athing')).slice(0,10).map(row => { const title = row.querySelector('.titleline > a'); const meta = row.nextElementSibling; return { title: title?.innerText, url: title?.href, points: meta?.querySelector('.score')?.innerText, comments: meta?.querySelectorAll('a')[meta?.querySelectorAll('a').length-1]?.innerText }; })") 

new Promise(r => {   const check = () => {     const el = document.querySelector('.content-loaded');     if(el?.innerText?.length > 100) r(el.innerText.substring(0, 5000));     else setTimeout(check, 1500);   };   check(); }) 

document.querySelector('.load-more').click(); window.scrollTo(0, document.body.scrollHeight); document.querySelector('input[name="search"]').value = '关键词'; 

write_file(path="/opt/hermes-home/projects/ai-neican/research/hn-20260514.md", content="[正文]")