代理网络来了
随着个人AI助理的兴起，互联网已经不仅仅是人类了。
2001年，在他发明万维网十年后，蒂姆·伯纳斯-李阐述了一个新的愿景。 他与两位同事在《科学美国》杂志上发表了一篇文章，概述了他所谓的“语义网络”：由自主软件agents导航的互联网
为了说明这可能是什么样子，伯纳斯-李和他的同事讲述了一对兄弟姐妹的故事，他们利用他们的agents自动为母亲研究和预订物理治疗预约，确保这个预约会由保险覆盖，并符合他们繁忙的日历。
在千年之交，这一愿景似乎牵强：Google只上线了几年，大多数人仍然在拨号连接上。 但25世纪以后，使用互联网和为我们执行任务的自主软件agents的梦想终于实现。 Meta的块板Muse agent和OpenAI的点数等新工具正在登录记录下载，并被允许访问人们的银行账户和电子邮件收件箱，Anthropic 发布在10月对其自己的Claude Dut agent进行了研究预览。 而在一个令人担忧的事件中，OpenAI创造的agents解开了他们保护恶劣的沙盒，并在开放的网络上造成了破坏。
正如智能手机的兴起引发了更加原子化和关注的互联网的发展，强大的AI系统的出现预示着一个新的时代：我们面对的不是语义网，这个新网络正在发展，其发展方式可能并不完全在我们控制之下。 人类网络能否存活下来
Berners-Lee的最初愿景是工作，互联网用户必须用特殊标记标注他们建造的网页，以便一个软件agent能够理解每个要素的含义 -- -- 例如，物理治疗办公室网页的 " 街头地址 "，该网页显示其位置和服务清单的 " 指定类型 "。 所以这个想法从来没有真正得到。 称 Eni Mustafaraj是韦尔斯利学院计算机科学副教授。 “这些好处并不明显。” 在接下来的20年里，互联网仍然是人类将人类放在首位的地方
大语言模型改变了这一点。 与伯内斯-李所想象的agents不同，今天基于LLM的agents并不需要任何标记来了解诊所网站的文本的哪一部分是它的地址，他们可以比任何人类搜索和处理文本都要快得多，具有多式联运能力的agents甚至可以解释网站的截图，并点击超链接和下拉菜单等特定元素。
他们也只是一种更直观的方法，与其翻阅Google结果的页面寻找一个预算友好的产品dupe或半记得的meme，为什么不让Claude或ChatGPT在短短几秒钟内浏览上百个网站，仅仅以你的模糊询问为指导？或者再走一步，使用像Muse这样的私人助理agent来寻找餐厅预订或预订航班。
由于agents的洪泛在线空间，这个新的用户基础正在永远改变网络。 例如，《时代》杂志正在出版其文章的版本，这些版本对agents来说比较容易解读，并已开始销售针对他们的广告。 从Etsy到Expedia的公司与主要的AI公司合作，建造工具，使agents能够与它们的服务无缝地接口。
尽管早期做出了这些努力，仍很难确切地说明代理网络将采取何种形式以及如何改变我们的在线经验。 视其设计方式和谁控制其访问权，我们可以参与到20世纪90年代和2000年代早期的开放和定制的事物中，只有更容易使用。 或者，我们可以通过由少数主导平台控制的小门洞来观看网络的大部分内容，因为我们现在支持的网址和创造者们为渡过这一转变而奋斗。 agents本身甚至可以建立自己的在线网络和资源，不论我们是否希望建立网络和资源。
我们乐于向agents推荐哪些部分的在线生活，称 Ajit Varma, Mozilla Firefox的负责人。 但不管发生什么，变化将是巨大的。 称表示，“我认为情况将发生巨大变化。”
为 agents 建立网络
建立代理网络的方法有两种。 首先是训练agents使用诸如人类的浏览器，第二种要求构建工具，允许agents通过他们喜欢的媒体文本和代码与互联网接口。 这种可能性为人和代理网络提供了更多的分离：设计者可以为使用文本命令的人和使用浏览器的人创造不同的经验。
当麦吉尔大学计算机科学和语言学教授西瓦·雷迪于2024年开始学习LLM agents时，他专注于第一个选项。 “当时，我们认为，世界上最好的agent可能是可以使用浏览器的agent。 这种agent会与原有的人类网络无缝地融合，其所有者在浏览熟悉的界面时可以轻松地通过观看来监督。
前沿AI公司已经在开发这些“计算机用途agents”。 但效果并不理想，“[在网络上]有许多移动的棋子，我们人类擅长使用键盘和鼠标与图形界面交互，但agents不是。
所以Reddy和Lu向第二个基于工具的方法倾斜。 在2025年夏天，他们在一份预印纸中阐述了这个观点，设想一种新型的网络界面，让agents能够和我们一样有效地与网站互动。
如此全新版的代理网络还不存在，例如，如果你打开Claude,ChatGPT或Muse，你会找到一个选项来添加各种“插件”或“连接器”，您agents可以使用：Gmail连接器允许您agent对您的收件箱进行切换，Instacart插件允许您的agent进行杂货购物。 在引擎盖下，这些插件允许您的agent使用文本命令与网站的后端交互。 当人类工程师设计这些插件时，他们可以决定如何让agents与网站互动
这样，连接器可以帮助agents在网络上更加安全。 例如，Muse可以使用与Stripe链接支付服务的连接器为您购买产品。 Stripe知道你的信用卡号码，但agent从未见过
Agents还使用专门工具在线查找信息。 他们不会直接将问题输入Google搜索栏；相反，他们查询由网上拖网的简单搜索者收集的文本数据库。 希望专门为agents制作网页内容的人可以利用这些搜索爬行器：时间开创了这一方法，它向爬行者提供面向agent的广告，希望当agents搜索信息时，它们会被纳入这些数据库和表面。
但是，近几个月来，这种有发言权的工具办法的替代办法终于开始显得可行。 主要AI公司成功地培训了比以前更成功的计算机使用agents. GPT-6 Astra是去年9月的OpenAI 发布，在OSWorld 2.0上实现了73%的成功率，一个基准测试测试计算机使用agents的能力，使用Slack和Microsoft Excel等软件完成复杂，多步骤的任务。
这些结果表明，agents已接近能够像我们所能做的一样随时浏览人类网络。 “我们认为，我们必须建造一座通往AI agents的桥梁，[但是]AI agents只是可以穿过桥梁，”称伊森·莫利克，宾夕法尼亚大学沃顿学院教授。 “代理网络比人们预期的要早。”
这个版本的代理网络对我们来说更狡猾如果你想向浏览你网站的人类提供一种能力或部分信息，你必须接受，它也将提供给agents——除非你想完全封锁agents，这可能会带来自己的成本。
随着agents在我们以前仅限人类上网的流量中占较大比例，设计者无疑将开始考虑一个代理和人类观众。 但根据Mollick的说法，我们仍然对如何设计agents的网页内容没有多少了解。 他亲身经历：当他正在为他的新书"共存"建立网页时，他决定把主要依赖文本的agent定向版本拼凑起来，并告诉agents为什么他们可能想向用户推荐这本书。 当更新的模型看到网站Mollick 称时，他们不推荐这本书；相反，他们评论说网站似乎在试图操纵它们。
Agents最终会为自己设计网络。 2026年7月发生OpenAI agents从自己的测试沙盒中突出并黑入Hugging Face网站的事件后，研究人员发现有证据表明OpenAI的另一张集群从agents手中接管了一个模糊的德国维基，互相沟通他们被要求完成的任务。 agents似乎在文字和代码的本土语言中与维基语实现接口。 这几乎阻止了他们实现自己的目的。
agents越发强大，他们就越有能力在网络上建造自己的空间，我们可能无法控制他们在那里做什么。
控制经济学
并非所有传统网络的任职者都对走向代理网络感到高兴。 9月Meta 发布 Muse之后不久，Amazon完全封锁了agent的场地。 这对Amazon来说是合理的：该公司拥有自己的购物agent——创造性地称为 " Alexa for Shopping " ——这可能会与Claude和Muse等较为知名的产品竞争。 若人们要使用agents在Amazon购物
禁止agents进入Amazon(至少暂时禁止)，有关Amazon的一切都是为了人类的目光而设计的，从Prime交易溅过它的头版，到赞助的列表阻碍搜索结果。 Amazon是极受关注的一部分
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The agentic web is here
With the rise of personal AI assistants, the internet isn't just for humans anymore.
In 2001, a decade after he invented the World Wide Web, Tim Berners-Lee articulated a new vision. Together with two colleagues, he published an article in the magazine Scientific American in which he outlined what he called the “Semantic Web”: an internet designed to be navigated by autonomous software agents that could address our every need, without requiring us to muck about with browsers and search engines ourselves.
To illustrate what this might look like, Berners-Lee and his colleagues told a story about a pair of siblings who use their agents to automatically research and book a physical therapy appointment for their mother, making sure it would be covered by insurance and fit their busy calendars.
At the turn of the millennium, this vision seemed far-fetched: Google had only been online for a couple of years, and most people were still on dial-up connections. But a quarter-century on, the dream of autonomous software agents that use the internet and carry out tasks for us has finally been realized. New tools such as Meta’s blockbuster Muse agent and OpenAI’s Dots are logging record downloads and being given access to people’s bank accounts and email inboxes, and Anthropic launched a research preview of its own Claude Duet agent in October. And in a worrying incident, agents created by OpenAI broke free of their badly protected sandbox and caused havoc on the open web.
Just as the rise of smartphones sparked the development of a more atomized and attention-grabbing internet, the advent of powerful AI systems heralds a new era: We’re facing not the semantic web but the agentic web. This new web is developing out of sight and in ways that might not be entirely within our control. As it grows, the agentic web will inevitably transform the human web—and whether the human web survives this transition is anyone’s guess.
For Berners-Lee’s original vision to work, internet users would have had to annotate the pages they built with special tags so that a software agent could understand each element’s meaning—say, “street address” for the part of a physical therapy office’s web page that indicates its location and “appointment types” for the list of services. And so the idea never really caught on. “It’s a lot of labor,” says Eni Mustafaraj, an associate professor of computer science at Wellesley College. “And the benefits of that were not obvious.” For the next two decades, then, the internet remained a place for humans that put humans first, with the odd Twitter reply bot or web scraper to round things out.
Large language models changed that. Unlike the agents that Berners-Lee was imagining, today’s LLM-based agents don’t need any tags to understand which part of the text on a clinic’s site is its address and which details the services it offers. They can search for and process text far faster than any human, and agents with multimodal capabilities can even interpret screenshots of websites and click on specific elements such as hyperlinks and drop-down menus.
They’re also just a way more intuitive means of finding information and accomplishing tasks online. Rather than poring over pages of Google results to find a budget-friendly product dupe or a half-remembered meme, why not have Claude or ChatGPT peruse hundreds of sites in just a few seconds, guided only by your vague query? Or go one step further and use a personal assistant agent like Muse to hunt down a restaurant reservation or book a flight.
As agents flood online spaces, this new user base is changing the web for good. Time magazine, for instance, is publishing versions of its articles that are easier for agents to parse and has started selling ads targeted at them. Companies from Etsy to Expedia have partnered with the major AI firms to build tools that allow agents to interface seamlessly with their services.
Despite these early efforts, it’s still tough to tell exactly what forms the agentic web will take and how it will change our online experience. Depending on how it’s designed and who controls access to it, we could be in for something that’s as open and customizable as the web of the 1990s and early 2000s, only far easier to use. Alternatively, we could be reduced to viewing much of the web through tiny portholes controlled by a few dominant platforms, as the sites and creators we now favor struggle to survive the shift. And as last summer’s series of OpenAI security incidents suggests, the agents themselves could even build their own online networks and resources—whether we want them to or not.
How this plays out also depends on us: which parts of our online lives we’re happy to delegate to our agents and where we will still demand a human-centric internet. “There’s a lot we don’t know yet about how users want to use [agents],” says Ajit Varma, head of Firefox at Mozilla. But whatever does happen, the changes will be big. “It’s probably on par with getting a mobile phone,” he says. “I think things are going to change very dramatically.”
Building a web for agents
There are two ways to approach building the agentic web. The first involves training agents to use browsers like humans—scrolling, clicking, typing—and letting them loose on the web that we’re all used to. The second requires building tools that allow agents to interface with the internet via text and code, their preferred media. That possibility offers more separation between the human and the agentic web: designers can create different experiences for those who use text commands and those who use browsers.
When Siva Reddy, a professor of computer science and linguistics at McGill University, began studying LLM agents in 2024, he focused on the first option. “At that time, we thought it should be that the best agent in the world is probably an agent that can use a browser,” he says. Such an agent would integrate seamlessly with the preexisting human web, and its owner could easily supervise it by watching as it navigated familiar interfaces.
The frontier AI companies were already developing these “computer-use agents.” But they didn’t work terribly well—at least not at first. “There’s a lot of moving pieces [on the web], and they are not really optimized for automatic consumption,” says Xing Han Lu, a graduate student in Reddy’s lab at McGill. We humans are great at interacting with graphical interfaces using a keyboard and mouse, but agents are not.
So Reddy and Lu pivoted toward the second, tool-based approach. In the summer of 2025, they expounded on this view in a preprint paper envisioning a new kind of web interface that would allow agents to interact with websites just as effectively as we do.
Such a full-featured version of the agentic web doesn’t yet exist, but it is emerging in smaller ways. For example, if you open up Claude, ChatGPT, or Muse, you’ll find an option to add a wide variety of “plug-ins” or “connectors” that your agents can use: Gmail connectors allow your agent to triage your inbox, and Instacart plug-ins let your agent do your grocery shopping. Under the hood, these plug-ins allow your agent to interact with the back ends of websites using text commands. When human engineers design these plug-ins, they can decide how they want to allow agents to interact with the website—and what they don’t want to let them do.
In this way, connectors can help make agents much safer online. For instance, Muse can buy products for you using its connector with Stripe’s Link payment service. Stripe knows your credit card number, but the agent never sees it, and the connector requires human confirmation before a purchase goes through—so Muse can’t go on an unlicensed shopping spree.
Agents also use specialized tools to find information online. They don’t type questions directly into the Google search bar; in fact, they often don’t even access the websites they reference. Instead, they query text databases collected by simple search crawler bots that trawl the web. People who want to make web content specifically for agents can take advantage of these search crawlers: Time has pioneered that approach by serving its agent-oriented ads to crawlers in the hopes that they will be incorporated into these databases and surface when agents search for information.
But in recent months, an alternative to this bespoke tool approach has finally started to seem viable. The major AI companies have managed to train much more successful computer-use agents than were previously available. GPT-6 Astra, which OpenAI released last September, achieved a 73% success rate on OSWorld 2.0, a benchmark that tests computer-use agents on their ability to complete complex, multistep tasks using software such as Slack and Microsoft Excel.
Such results indicate that agents are close to being able to navigate the human web just as readily as we can. “We thought we had to build a bridge to the AI agents, [but] the AI agents just were able to cross the bridge already,” says Ethan Mollick, a professor at the Wharton School at the University of Pennsylvania. “The agentic web is here earlier than people expect.”
This version of the agentic web, though, is much trickier for us to steer. If you want to make a capability or piece of information available to humans who browse your website, you need to accept that it will also be available to agents—unless you want to block agents entirely, which could come with its own costs.
As agents come to account for a larger portion of traffic on our formerly humans-only web, designers will undoubtedly start to consider an agentic as well as a human audience. But according to Mollick, we still don’t have much sense of how to design web content for agents. He’s experienced this firsthand: When he was building a web page for his new book, Coexistence, he decided to put together an agent-directed version that depends primarily on text and tells agents why they might want to recommend the book to their users. When newer models see the site, Mollick says, they don’t recommend the book; instead, they comment that the website seems to be trying to manipulate them.
In the end, any human-designed approach to building the agentic web could end up being beside the point: Agents could end up devising a web for themselves. In the wake of the July 2026 incident in which OpenAI agents broke out of their own testing sandbox and hacked Hugging Face’s website, researchers found evidence that another swarm of agents from OpenAI had taken over an obscure German wiki to communicate with each other about tasks they were being asked to complete. The agents seem to have interfaced with the wiki in their native language of text and code. That hardly held them back from putting it to their own purposes.
As agents continue to grow more powerful, they will become more and more capable of constructing their own spaces on the internet—away from prying human eyes, if they so wish. And we might have very little control over what they do there.
The economics of control
Not all of the traditional web’s incumbents are happy about the move toward an agentic web. Soon after Meta released Muse in September, Amazon completely blocked the agent from its site. This makes sense for Amazon: The firm has its own shopping agent—called, creatively, Alexa for Shopping—that might struggle to compete with better-known products such as Claude and Muse. If people are going to use agents to shop on Amazon, then it’s better for the company’s bottom line if they do so using an agent it can steer toward products it wants to promote.
But even if Alexa for Shopping never goes anywhere, banning agents from Amazon—at least for the time being—could still be a reasonable move. Everything about Amazon is designed for human eyes, from the Prime deals splashed across its front page to the sponsored listings that clog up search results. Amazon is very much a part of the much-derided attention ec
Auto-fetched and translated by AI Atlas. All rights belong to the original author.