HackerNews中文版

这些旧方式太繁琐了。完全自我浏览并不需要埃隆·马斯克的视觉处理能力。它只需要一种“代理语言”——代理的母语，LLM（大语言模型）的通用语——简单的纯文本。还有诚实。关于它能做什么（网络上的一切，除了只有你能做的事情），以及它不能做的事情，但你能做的：多因素认证（MFA）、验证码、登录。一个具备智能护栏的代理技能和一个设计良好的Unix哲学命令行工具足以完成网络上的任何任务。你也可以试试。以下是我给它提出的一些任务，它都完成了：- 找到至少100条相关的推文，并撰写适当的回复，向那些面临痛点的人推广WebCLI Grok Build with Composer 2.5 快速- 比较从旧金山（SFO）到华盛顿特区（DC）下午的航班，使用谷歌航班、Kayak等多个供应商，找到最便宜的直达航班。填写我的信息并预订，但在付款时停止。 Codex 5.5 高- 在亚马逊、沃尔玛、阿里巴巴和lego.com上找到一些有趣的乐高产品，寻找最酷的套装或大数量的积木，以最佳价格找到，并一路到结账，填写详细信息。 Claude Sonnet 4.6还有更多。代理总是能找到解决方案。没有截图，没有选择器，只有原始文本和带有编号的操作参考，诚实地进行验证，并提供一系列有用的表面信息。核心循环是一个简单的OODA循环：<pre><code> web inspect # 代理观察和定位 web do <ref> <opts> # 代理决定并行动 </code></pre> 重复。永远。这能完成所有任务。这个工具足够小，足够透明，代理足够聪明，足够坚持，总能找到解决方案。这是一个智能驱动的网络任务新纪元。再也没有剧作家和木偶，也没有机械的“自动化”哔哔声。WebCLI是由代理智能驱动的网络即兴创作。我想继续为代理能力构建技术。想象一下，如果你试用这个工具，它能为你节省时间和繁琐的工作。试试看，然后因为它的价值付费。每个电子邮件域名可以免费获得一个功能齐全的五天试用，只需提供电子邮件。Https://webcli.sh如果你有想法希望大规模使用它，请联系我。

查看原文

These old ways are too heavy. Full self browsing doesn’t require Elon Musk vision processing.It just requires Agentish - the agent’s native tongue, the LLM’s lingua franca - frickin plain text.And honesty. About what it can do (everything on the web, besides stuff only you can do), and what it can’t do, but you can: MFA, captcha, login.An agent skill with smart guardrails and a well designed Unix philosophy CLI tool is enough to power any task on the web.You can try it too. Here’s some things I’ve thrown at it and it’s done:- find at least 100 relevant tweets and craft apt replies that promote WebCLI to people experiencing the pain it solves Grok Build with Composer 2.5 Fast- Compare flights from SFO to DC mid afternoon across a couple of providers like Google flights and kayak, etc.. Find the cheapest one with no stops. Fill in my details and book it but stop at payment. Codex 5.5 high- Find some fun Lego products across amazon Walmart Alibaba and lego.com and find the coolest set or large quantity of blocks at the best price and take it all the way to checkout filling in details. Claude Sonnet 4.6And many more. The agent’s always figure it out. No screenshots no selectors, just raw text and numbered references for actions with honest in validation and a bunch of useful surface.The core loop is a simple OODA loop:<pre><code> web inspect # agent observes and orients web do <ref> <opts> # agent decides and acts </code></pre> Repeat. Forever. That gets everything done. The tool is small enough, and transparent enough, and agents are smart enough, and persistent enough that they always figure it out.It’s a new era of web task driving with intelligence. No more playwrights and puppets, no more robotic “auto-mation” beep boop. WebCLI is web improvisation, powered by agents’ intelligence.I want to keep building technology for agency. Imagine if you tried this tool and it saved you time and drudgery. Try it and then pay because it’s valuable. You get a free, fully functional five day trial per email domain, with just email. Https://webcli.shContact me if you have ideas to use it at scale

告诉HN：忘掉选择器和截图吧。自主网络存在于你的终端中。