E-commerce websites might also try to pivot to the agentic attention economy, but the path forward there is unclear. Last year, a team led by Akshit Kumar, an assistant professor of operations management at the University of Toronto, investigated whether and how agents are influenced by the ways that shopping websites present search results. They found that agents entirely ignore items that come with a “sponsored” tag, which suggests that selling favorable positions on results pages might not be a great way for sites to make money in an agentic economy. Models did seem to have a preference for items that appeared in particular positions in search results—but whether they prioritized items at the top, bottom, or middle of the results page differed from model to model and between generations of a single model. It could be tricky for e-commerce platforms to keep up.
David Rothschild, an economist at Microsoft, says he’s hopeful that the transition to the agentic web could put more power in the hands of the consumer. Paying for online content today typically requires a commitment: We subscribe to a magazine or join a YouTuber’s Patreon, and in turn we gain access to their full output. One alternative is micropayments that let users buy content à la carte, but the idea has foundered in the past because paying article by article or video by video requires too much administrative overhead. Agents, though, could take care of that overhead for us, letting us pay directly for each piece of content we consume and putting just a bit of cash directly in the hands of each person whose work we enjoy.
Other experts share a similar vision. Mustafaraj, the Wellesley computer scientist, anticipates that agents could help us escape from the algorithms that control so much of the media we consume. “My hope is that we can have AIs that work for us, in the sense that they can change the feed the way I want—not the way TikTok wants,” she says. And Varma, of Mozilla, imagines an internet without intermediaries like Amazon and Doordash, which now extract tolls for facilitating interactions between consumers and businesses. Instead, he imagines, perhaps we will all have agents that can securely and reliably interface directly with, say, restaurants’ agents to get dinner delivered to us, free of pesky service fees.
It’s an appealing idea, but the history of the web has been a story of consolidation and concentration of power and control, and that might be the default path here, too. When OpenClaw, the first popular personal assistant agent, exploded onto the scene in early 2026, it seemed, for a brief moment, that agents might be built in a grassroots, distributed way. But OpenAI quickly snapped up OpenClaw’s creator, Peter Steinberger, and other frontier AI companies got to work building their own personal assistant agents.
When we use those agents, we’re at the mercy of how the companies want their tools to behave. “You have this technology that can just eliminate all this communication friction,” Rothschild, the Microsoft economist, says. “You’re going to see companies build barriers in order to maintain the consumer base that they have.”
Here’s what that could look like: Let’s say Muse becomes the default way that many people find information, shop, and manage their digital lives. In principle, users could deploy Muse to scour the internet for the content most valuable to them, connect directly with vendors, and otherwise escape the walled gardens that have defined the internet for the past couple of decades. But Meta might not want to let them do that. Instead, it might design Muse to keep users within its own walled garden—by, say, preferentially showing users content from Meta social networks and recommending businesses that have paid a fee to be featured prominently in Muse results.
There are some benefits to this more closed-off version of the agentic web. Agents can be powerful tools for hackers and scammers. The major companies have safeguards in place to prevent, or at least mitigate, this misuse—for example, their AI tools automatically identify when someone is making a hacking-related request and refuse to respond. In a fully open ecosystem, however, someone would surely build and disseminate a system with no such restrictions. Companies also have a strong reputational incentive to prevent their personal assistant agents from making major mistakes, such as authorizing a large purchase that the user didn’t want—and so they’ve put plenty of effort into building reliable guardrails.
