MIT Technology Review published a major overview, 'The agentic web is here' (by Grace Huckins, November 2, 2026), with the central thesis that the internet is becoming 'agentic'—increasingly used not by humans but by autonomous AI agents that read email, pay for purchases, and make bookings on a person's behalf. This practically realizes Tim Berners-Lee's long-standing idea of the Semantic Web, but through LLMs rather than page markup. The overview shows that this is already an economic shift, not just a technological one: agents ignore advertising, search referrals are dropping, and major platforms have begun blocking third-party agents in favor of their own.
What Happened
The overview describes the agentic internet as an already established technical shift with two working architectures. The first path is computer-use agents that control a browser like a human: the GPT-6 Astra model, released by OpenAI in September 2026, showed a 73% success rate on the OSWorld 2.0 benchmark. The second path is tool-based: agents act through connectors and plugins to services, including Gmail, Instacart, and Stripe Link, as named in the overview. Simultaneously, the material records the first measurable consequences: Amazon blocked Meta's Muse agent to promote its own Alexa for Shopping; a University of Toronto study showed that agents completely ignore products labeled 'sponsored'; and Pew recorded a drop in referrals to websites of approximately 40% against the backdrop of the spread of AI Overviews.
Context
The idea of a machine-understandable internet is not new: in 2001, Tim Berners-Lee described the Semantic Web in Scientific American, proposing to make the web accessible to machines through standardized page markup. This project was not realized at the time, but today large language models solve the same task: an agent interprets a standard 'human' interface, and the web's structure is built through learning rather than data formats. The two approaches to the agentic web are complementary rather than competing: browser control is universal and works with any interface, but is fragile and has a broader risk surface, while connectors are limited in service coverage but allow for strict limitation of agent permissions. As a result, the interface to services ceases to be exclusively human: in parallel with websites and apps, a machine access channel is growing, where the decisive factor becomes the question of trust and the scope of delegated rights.
Why This Matters for the Industry
The attention economy—advertising, subscriptions, marketplaces—loses its foundation when the 'user' becomes an agent: machine clients do not create impressions or clicks and do not react to targeting, because the policy of their decisions is built on different premises than human attention. Publishers and online stores will have to choose between blocking third-party agents and building separate interfaces for them, and the overview describes new monetization mechanisms: agentic advertising formats (the Time case), micropayments for publishers, and direct connectors to Stripe and Instacart. Some markets will go the path of 'walls,' and then distribution will migrate to the connector layer of major agentic ecosystems, where owners of trusted connectors gain control over access to the audience.
Why This Matters for Users
For the reader, this is primarily a question of the security of their own access rights: agents like Meta Muse, OpenAI Dots, and Claude Duet already work with email, bank accounts, and bookings, so it is important to understand which services you trust an agent with and where human confirmation of actions is enabled. The Stripe Link pattern shows a working solution: an agent can be entrusted with payment without showing it the card number. A separate aspect is accessibility: for elderly people and people with disabilities, an agent becomes a tool that removes routine operations from complex interfaces. The flip side is that part of the 'human' web may become less accessible for humans: bookings, tickets, and advertising platforms will be restructured for machine clients.
What Is Still Unknown / Limitations
Key figures from the overview should be read with caution. The 73% result on OSWorld 2.0 is a vendor figure from a single journalistic piece: the sources do not describe the methodology, the definition of successful task completion, the cost of tasks, the number of steps, or latency, and there is no independent replication. Furthermore, this is a task-level success metric, not a step-level one: step-by-step reliability is a different quantity, not disclosed in the material, and in long multi-step scenarios, failures accumulate according to their own laws, so the interpretation of '27% failure at each step' is incorrect. The University of Toronto and Pew data are single studies, and conclusions about the restructuring of the attention economy remain interpretations rather than confirmed facts.
Sources
Author
Look at AI, editorial team