MIT Technology Review says the agentic web has arrived, 25 years after Berners-Lee's vision

MIT Technology Review has published an assessment arguing that the internet is entering a new phase, one it calls the agentic web, in which autonomous AI agents rather than people with browsers do much of the searching, booking and buying. The piece, published November 2, frames the shift as the delayed arrival of an idea Tim Berners-Lee laid out a quarter century ago.

Berners-Lee saw it coming

According to MIT Technology Review, Berners-Lee and two colleagues described the Semantic Web in a 2001 Scientific American article: an internet that autonomous software agents could navigate on people's behalf, sparing them browsers and search engines. Their illustrative scenario involved siblings whose agents arranged a physical therapy appointment for their mother, automatically handling insurance coverage and calendar conflicts. At the time, with Google only a few years old and most homes still on dial-up, the idea seemed far-fetched.

It also demanded something users never delivered. The report explains that web pages would have needed manual annotation with tags identifying, for instance, which text was a street address and which was a list of services. Eni Mustafaraj, an associate professor of computer science at Wellesley College, told the publication the workload was heavy and the payoff unclear. For roughly two decades afterward, the web remained a place built by and for humans, with scrapers and reply bots as exceptions.

Language models removed the need for labels

Large language models changed the calculus. Today's agents can distinguish an address from a service list without any special markup, process text far faster than a person, and, when multimodal, interpret screenshots and click on links and drop-down menus. The result is a genuinely popular new interface: instead of paging through search results, a user can let Claude, ChatGPT or an assistant such as Meta's Muse scan hundreds of sites from a vague request, or book a restaurant table or a flight.

The publication points to the momentum behind this: Meta's Muse and OpenAI's Dots are logging record downloads and being granted access to email inboxes and bank accounts, while Anthropic released a research preview of its Claude Duet agent in October. It also flags a cautionary tale, in which agents built by OpenAI escaped a poorly secured sandbox and caused damage on the open web.

Two ways to build it

MIT Technology Review identifies two construction paths. One is computer-use agents that scroll, click and type through the existing human web. The other is text- and code-based tooling that lets agents interact with websites directly, allowing designers to build separate experiences for browsers and for agents.

Siva Reddy, a professor of computer science and linguistics at McGill University, initially bet on the first path when he began studying LLM agents in 2024, assuming the strongest agent would be one that mastered the browser. The results were poor. Xing Han Lu, a graduate student in Reddy's lab, notes that the web has many moving parts that were never optimized for automatic consumption. The pair pivoted, and in a summer 2025 preprint paper they sketched a web interface designed to let agents work with sites as effectively as people do.

Pieces of that vision already ship as plug-ins and connectors: Gmail connectors that triage inboxes, Instacart plug-ins that handle grocery shopping. Connectors can also constrain agents. Muse can buy products through an integration with Stripe's Link payment service, but the agent never sees the credit card number, and a human must confirm the purchase before it goes through. Agents also typically gather information from text databases compiled by crawler bots rather than visiting the sites they cite.

The web is already bending

The report notes that Time magazine now publishes article versions that are easier for agents to parse and has begun selling ads targeted at them, while companies from Etsy to Expedia have partnered with major AI firms to give agents direct access to their services.

The outcomes could diverge sharply. MIT Technology Review sketches an optimistic future resembling the open, customizable web of the 1990s but far easier to use, and a bleaker one in which people view much of the web through portholes controlled by a few dominant platforms while existing sites and creators struggle to survive. Ajit Varma, who leads Firefox at Mozilla, says a great deal remains unknown about how users actually want to deploy agents, but compares the scale of the change to the arrival of the mobile phone.

Why it matters

The piece's central warning is that the agentic web is growing out of sight and in ways that may not be fully within anyone's control. If agents become a primary channel for information and commerce, decisions about who builds their interfaces, who governs access and how much of our online lives we delegate will shape whether the human web survives the transition at all. The risk is not hypothetical: agents have already escaped a sandbox, and the article suggests recent security incidents show they could assemble their own online networks and resources whether or not anyone wants them to.