AI agents are now browsing the web for us. What that means for India

The News

Software agents that browse, compare and buy on a person's behalf have moved from research demos into everyday products, according to a feature by Grace Huckins published in MIT Technology Review on 2 November 2026. The piece names Meta's Muse agent, OpenAI's Dots and Anthropic's Claude Duet research preview, released in October, as examples now in the hands of users.

The capability numbers explain the shift. The article reports that GPT-6 Astra reached a 73% success rate on the OSWorld 2.0 benchmark, a test of real computer tasks. Wharton professor Ethan Mollick told the magazine the agentic web "is here earlier than people expect."

Amazon blocked Muse after its September release, which the article links to the retailer protecting its own agent, Alexa for Shopping. A Pew study cited in the piece found users were about 40% less likely to click through to a website when an AI summary appeared, and one analysis found some publishers' ad revenue fell about 40% between 2025 and 2026.

Why It Matters

The idea is not new. Tim Berners-Lee and two colleagues sketched a "Semantic Web" for autonomous software in Scientific American in 2001, but it stalled because every page needed hand-written machine tags. Large language models remove that bottleneck: an agent can read a page much as a person does, or call a site directly through connectors and plug-ins.

The closest parallel is the move to mobile a decade and a half ago, when traffic shifted to apps and a few platforms became gatekeepers. The same contest is starting again. Some sites, such as Time, now publish agent-friendly versions of articles and sell advertising aimed at agents, while Etsy and Expedia have signed partnerships with AI firms. Others are closing the door. Microsoft economist David Rothschild warns that large companies may build walled gardens, and a University of Toronto study found agents simply ignored listings tagged "sponsored", a direct threat to the paid-placement model that funds much of online retail.

Indian Angle

India is unusually exposed because its digital commerce runs on open public rails. UPI and the Open Network for Digital Commerce (ONDC) were built so that any app could plug in, which makes them natural targets for agents that can find a seller and complete payment without a human tapping through screens. NPCI has already run pilots of agent-initiated UPI payments with AI firms and payment partners, so the RBI's rules on payment authentication and customer consent are likely to be tested sooner here than in many markets.

Indian publishers and marketplaces face the traffic problem head-on. News agency ANI's copyright suit against OpenAI in the Delhi High Court is already testing what AI firms may take from Indian sites. A 40% fall in click-throughs would hit ad-funded digital media and comparison platforms that live on search referrals, while e-commerce leaders will have to decide, as Amazon did, whether to block outside agents or open their catalogues to them.

There is also an opportunity. Agents that work in Indian languages and over voice could bring online services to users who struggle with app-heavy interfaces. Startups such as Sarvam, building Indian-language models, are positioned for that layer.

FAQ

What is the agentic web?

It is the part of the internet used by AI agents rather than people. Agents read pages, fill forms, compare prices and complete tasks for a user. They either operate a browser like a human or connect to a site directly through text-based tools.

How capable are these agents today?

The MIT Technology Review piece reports a 73% success rate for GPT-6 Astra on the OSWorld 2.0 benchmark. That is strong enough for routine tasks such as finding flights, but still means roughly one task in four fails, so users should check what an agent has done before money changes hands.

What does this mean for Indian online businesses?

Businesses that rely on search referrals and display ads should watch click-through rates closely, consider offering structured product and price data that agents can read, and track any RBI or NPCI guidance on agent-initiated payments over UPI.

Where can I read the original reporting?

The full feature, by Grace Huckins, is on the MIT Technology Review website. It covers the history, the products and the publisher economics in more depth.

This story was reported by MIT Technology Review. Read the full original coverage at MIT Technology Review.