What Is the Agentic Web? The Impact of an Era Where AI Uses the Internet
機械翻訳 / Machine-translated
A
AI Friends
1日前・19分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Wouldn't it be convenient if AI could search, compare, and even complete payment on your behalf when you shop online? That future is already becoming reality. Based on reporting from MIT Technology Review, this article breaks down how the agentic web works and what it means for operators and users.
In an article published November 2, 2026, MIT Technology Review declared, "The agentic web is here."
The agentic web refers to a state in which AI agents—AIs that navigate the web and carry out tasks on behalf of users—use the internet alongside humans.
The question the article poses is whether a human-centric web can survive this shift.
A similar concept actually existed back in 2001. Tim Berners-Lee, the inventor of the Web, and colleagues published a piece on the "Semantic Web" in a scientific journal—a world in which software would browse the web in place of humans.
That vision never took hold, however. Site operators would have had to attach special tags (markers) for machines to read on every single page. Eni Mustafaraj of Wellesley College explains that the effort was enormous while the benefits were hard to see.
This time, things are different. Agents powered by LLMs (AIs that handle language much like humans do) can read ordinary text without any tags. They can also look at images of screens and click links and menus.
The article mentions Meta's Muse, OpenAI's Dots, and Anthropic's "Claude Duet" as personal agents capable of accessing email and bank accounts.
Time magazine has prepared article versions for agents and is also selling ads targeting agents. Etsy and Expedia have partnered with major AI companies to create dedicated gateways for agents.
In terms of capability, the article reports that OpenAI's GPT-6 Astra achieved a 73% success rate on OSWorld 2.0, a benchmark measuring multi-step software operations. Ethan Mollick of Wharton says the agentic web "arrived faster than people expected."
There are broadly two ways agents interact with the web.
The first is opening a browser and performing clicks and inputs—the same screen humans use, which makes it easier to work with virtually any site.
However, Shiva Reddy and Xin Han Lu of McGill University note that early performance with this method was not strong.
The second method involves communicating through text or code. Using connectors (components that link services together) or plugins, agents work without ever looking at a screen.
A paper published by Reddy and colleagues in June 2025 argues that "the web should be built for agents"—the idea of stopping the practice of adapting agents to human-facing screens and instead building new interfaces designed for agents.
Here are some concrete examples. Using Gmail's connector, AI can organize your inbox. With Instacart's plugin, grocery shopping can be delegated entirely.
When money is involved, security mechanisms are essential. According to the article, Muse can make purchases using Stripe's Link connector. In this process, the agent cannot see card numbers, and a human must confirm each purchase.
Stripe's official documentation describes the same flow. The agent creates a "spend request" specifying the amount and merchant, and once the customer approves it, credentials valid only for that purchase are returned. The customer's actual card number is never exposed. Approval windows last 30 minutes.
Note that this payment feature is currently available only to customers in the United States and Canada.
Behind the convenience, conflict has already emerged. Amazon locked Meta's Muse out of its site.
Muse launched on September 8, 2026, and the lockout came roughly 12 days later. According to Adweek's reporting, users who tried to shop via Muse were shown a popup stating that continued access by unauthorized AI agents violates Amazon's terms of service.
Amazon cited three reasons: Meta gave no advance notice; Muse does not identify itself while browsing; and there are privacy and security concerns about how login credentials are handled.
Meta countered that Muse cannot see passwords or payment information. On this point, Amazon's account and Meta's claim remain in conflict.
Context: Amazon also sued Perplexity in November 2025, alleging that Perplexity's shopping agent Comet accessed accounts without permission and disguised automated operations as human browsing. Perplexity has denied this.
Amazon, meanwhile, operates its own agent, "Alexa for Shopping." The picture that emerges is less about shutting out agents entirely and more about controlling who passes through its own marketplace.
Much of the web runs on advertising, sponsored placements, and page view counts. When agents visit on behalf of humans, this model weakens.
The MIT Technology Review article cites several figures. One study found that when AI summaries appear in Google search results, users are about 40% less likely to click a link. One analysis found that some publishers' ad revenue fell by roughly one-third between 2025 and 2026.
So, are agents influenced by advertising? Research by Akshit Kumar and colleagues at the University of Toronto Scarborough found that agents were less likely to recommend products labeled "Sponsored." However, preferences for placement position varied by model and generation.
Mollick even ran an experiment, creating a dedicated page to get agents to recommend his book. Newer models didn't take the bait—one responded that "this site appears to be attempting manipulation."
Experts are divided.
On the optimistic side, David Rothschild of Microsoft Research says micropayments—paying a small amount per article—that were previously impractical might become possible. Mustafaraj sees potential in stepping away from platform algorithms to build information flows tailored to oneself.
The pessimistic concern is lock-in. Rothschild raises the possibility that, for example, Muse might prioritize its own content or paid merchant listings. This is a hypothetical example, and neither Meta nor OpenAI commented to the reporter.
Security presents difficult trade-offs as well. Closed systems make it easier to prevent hacking and costly mistakes, but at the cost of equitable access to growth.
The article draws a parallel to the iPhone's launch in 2007. Smartphones didn't replace the web, but some services became unusable without them. Similarly, a day may come when people who don't use agents are left at a disadvantage.
There are upsides too. Margaret Mitchell of Hugging Face points out that voice-driven agents could help elderly people who struggle with screen navigation, as well as people with visual or motor disabilities.
What about Japan? This article focuses primarily on developments in the United States. Data measuring the impact on the Japanese market was not found within the scope of research. What follows is a framework for thinking through the issues.
For those running e-commerce sites, it may be wise to write prices, stock levels, and specifications clearly in text, in anticipation of agents coming to read product pages. Pages that are easy for humans to read are also easy for agents to read.
Media operators need to prepare for a drop in clicks due to AI summaries. It may be time to consider value that AI summaries cannot replicate—such as membership models or proprietary data.
For users, what matters is deciding for yourself how much to delegate to an agent. Cristian Coloandra of Opera says he's happy to hand off defined tasks like comparing airfare, but wants to plan his own vacation.
Developing the habit of checking the amount and merchant on payment confirmation screens every time is something anyone can start now.
It refers to a state in which AI agents browse the web on behalf of users—comparing options, making purchases, and booking reservations. Humans and agents use the same web.
The Semantic Web required operators to attach special tags. Today's agents can read ordinary text and screens, so they can operate without tags.
With Stripe Link, agents cannot see actual card numbers, and owner approval is required for each purchase. However, this is currently available only to customers in the United States and Canada.
Amazon cited a lack of advance notice, Muse not identifying itself while browsing, and concerns about how login credentials are handled. Meta countered that Muse cannot see passwords or payment information.
First, write information such as prices and specifications clearly in text. It also helps to decide in advance whether to accept or reject agents.
Start by checking whether the information on your site is in a form that AI can read correctly.
This article is a cross-post from AI Friends.
