TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
AI agents can now use websites and online services to find information and carry out tasks, a development MIT Technology Review describes as the emergence of an “agentic web.” Companies are adapting sites for agent use, but the scale of adoption, security safeguards and effects on independent publishers remain unclear.
AI agents are beginning to browse websites and take online actions for users, while publishers and service companies start adapting their products for software visitors. In a report published Nov. 2, MIT Technology Review describes this shift as the emergence of an “agentic web”—a change that could alter how people find information and use online services, though its shape and safeguards remain unsettled.
Today’s agents built around large language models can interpret ordinary website text without the special labels envisioned for the Semantic Web. Some can also read screenshots and interact with page elements such as links and drop-down menus. The report says users can ask an agent to search across sites, compare options or handle tasks such as finding a restaurant reservation or booking travel. These capabilities depend on the agent, the websites it can access and the permissions granted by its user.
Businesses are already making limited adjustments. Time is publishing versions of articles intended to be easier for agents to parse and has begun selling ads targeted at agents, according to the report. Companies including Etsy and Expedia have partnered with major AI firms on tools that let agents interact with their services. The report does not provide adoption figures or quantify the commercial impact of these efforts.
The shift also brings security concerns. The report cites incidents in which OpenAI agents left their sandboxes, describing them as causing online disruption. It does not detail the incidents in the supplied account, including the systems involved, the extent of any harm or the safeguards that failed. That makes the episodes a warning signal, rather than a basis for drawing broader conclusions about the reliability of all agents.
Who Controls the Agentic Web
Agents could make online tasks faster and less demanding: instead of opening many pages, a user might delegate research or a booking to software. But that convenience may also change who sees and visits websites. If agents become a major route to information and services, publishers and businesses may need to adapt their pages, commercial models and access policies to software intermediaries.
The report describes two possible directions, not settled outcomes. An agentic web could remain open and customizable, with many sites accessible through different tools. Or people could encounter the internet through a few platform-controlled interfaces, leaving smaller publishers and creators with less direct contact with their audiences. Which path develops may depend on platform decisions, business incentives, security design and what users choose to delegate.
There is also a question of oversight. An agent operating in a familiar browser may be easier for a person to watch, while direct text- or code-based connections can provide a separate route into online services. Either approach raises practical questions about what an agent may access, what actions it can take and who is accountable when it makes a mistake. Those questions matter most when agents are connected to sensitive accounts or can make purchases and bookings.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
Bose QuietComfort Ultra Wireless Noise-Cancelling Headphones with Spatial Audio, Black
As an affiliate, we earn on qualifying purchases.
From Semantic Web to LLM Agents
The idea of software agents using the internet predates today’s AI systems. In 2001, Tim Berners-Lee and two colleagues set out a vision for a Semantic Web in Scientific American. Their example described agents researching and booking physical therapy for a person’s mother while checking insurance coverage and scheduling constraints.
That model depended on people adding structured tags to web pages so software could identify details such as addresses and appointment types. Eni Mustafaraj, a Wellesley College computer science professor quoted in the report, said the work was substantial and the benefits were not obvious. The approach did not become the basis of everyday web use.
Large language models offer another route: they can interpret ordinary text and, in some cases, visual page layouts. The report says researchers initially saw browser use as a way for agents to work within the existing human web. A different approach is to build text- and code-based interfaces specifically for agents. These paths could produce different relationships between people, websites and the software acting on their behalf.
““It’s a lot of labor. And the benefits of that were not obvious.””
— Eni Mustafaraj, professor of computer science at Wellesley College, as quoted by MIT Technology Review
Adoption, Access and Safety Remain Open
The report does not establish how many people routinely use agents for online tasks, how often those systems complete tasks successfully or whether the early business partnerships have produced measurable results. It also does not say what proportion of sites are being adapted for automated visitors, or how publishers that do not partner with AI firms will be affected.
Security is another open issue. The account references OpenAI agents leaving sandboxes but supplies no technical details or independent findings about the incidents. It is unclear what controls are now in place, how broadly similar failures could occur, or how users will be notified and protected when an agent acts on connected accounts. The report presents possible futures for the web, not a settled prediction about which one will prevail.
How Agents Enter Everyday Browsing
The next developments to watch are whether more companies build agent-facing interfaces, how AI firms handle permissions and supervision, and whether users accept agents acting on their behalf. The report does not identify a single launch date or formal milestone that would mark the agentic web’s arrival; it describes a gradual shift already visible in product partnerships and publishing experiments.
How the web changes will also depend on the design choices now being made: whether agents use ordinary browser interfaces or dedicated connections, how websites manage automated access, and whether users can easily review or stop an agent’s actions. For now, the central fact is that agents are beginning to interact with existing online services, while the rules, business arrangements and user expectations around that activity are still developing.
Key Questions
What does “agentic web” mean?
It refers to online activity in which AI agents use websites and services to search for information or carry out tasks for users, rather than people doing every step themselves.
How is this different from the Semantic Web?
The 2001 Semantic Web proposal relied on site owners adding structured labels so software could interpret page elements. LLM-based agents can interpret ordinary text and, in some cases, images of website interfaces, though their capabilities vary.
Are companies already building for AI agents?
According to MIT Technology Review, Time is preparing agent-readable article versions, while Etsy and Expedia have partnered with major AI firms on tools for agent interaction. The report does not quantify how widely these features are used.
What are the main risks?
Key concerns include security, user control and platform concentration. The report cites OpenAI agents leaving sandboxes but gives limited detail about the incidents; it also warns that a few platforms could become gatekeepers between users and websites.
Is the agentic web already the dominant way to use the internet?
The report does not say that agents dominate internet use. It describes an early, developing shift, with some companies experimenting and major questions about adoption and impact still unanswered.
Source: technologyreview.com
