In detail
Most of the web will never re-architect for agents
Dhruv Batra’s bet is that infrastructure changes slowly. The head of the distribution, the top 50, 100, 500 sites holding enterprise and customer data, has the incentive and maturity to build agent pathways or charge tolls for access. The long tail does not. Batra points to roughly 20,000 to 40,000 school district sites in the US, plus government offices and hundreds of thousands of e-commerce pages, each built for people. He gives a San Francisco tennis-court reservation system written by one indie developer as the pattern. It works, and it will not reinvent itself overnight. Some of the holdout, in his words, is socio-political, and pointing coding agents at these sites does not solve that part.
Pixels in, clicks out
Batra’s model for action skips the API question. “If a machine can perceive pixels and click buttons and do the things and report to you that yes, I’m able to do it. There, that’s your API.” The source of truth is what the human sees on screen, not the DOM or the accessibility tree. Batra says those internal structures are inconsistent, because there is no central owner of the web and no enforced standard. A browser agent looks at the rendered screen, decides which button to click, and acts. This is a different bet from tool-calling and MCP, which assume a site exposes structured endpoints. Here the human-visible screen is the interface, and no re-architecting is required. There is now a lighter middle path, such as Chrome’s WebMCP early preview, which lets a site describe its forms and actions to agents without building a separate API. Batra’s argument applies there too: it still asks each site to do the work.
The web is a shared roadway
The analogy Batra reaches for is autonomous driving. Cities could make self-driving easy by building dedicated lanes wired with sensors, but that infrastructure is too expensive to lay down. What exists instead is shared autonomy, with human drivers and autonomous cars on the same roads. Batra applies the same logic to the web. Thirty years of browsers and websites made rich visual information efficient for the human brain, and that is the roadway agents will share for a long time to come. Agentification happens through machines that act like people, perceiving pixels and clicking buttons, rather than through a web that rebuilds itself for machines.
Why this favors small specialized models
Batra argues the pixels-in workload rewards a specialized model over a frontier one. A browser agent runs in a perceive-act loop over long trajectories, so each step has to be cheap and fast or the task stops making economic sense. Yutori’s Navigator is post-trained in-house for browser and computer use. At the time of the episode, Batra said it was slightly more accurate than Opus 4.7 and GPT-5.5 on standard browser-use benchmarks, a gap he declined to make much of; the number he emphasized was efficiency: 2 to 3 times faster and 4 to 5 times cheaper, because it is small and light rather than general-purpose. Version N1.5 also writes JavaScript on the fly to shorten trajectories, filling several form fields in one action instead of clicking each field by hand.
Hear it from the guest
“If a machine can perceive pixels and click buttons and do the things and report to you that yes, I'm able to do it. There, that's your API.”
“The Golden Gate Park tennis court reservation system website was written by some indie developer. That website has a perfectly functioning calendar view reservation system. It's not going to reinvent itself overnight.”
“Instead of trying to reinvent every part of the software stack, the Frontier Labs just made it really easy to operate on top of existing tools everyone's already using.”
Quotes lightly edited to remove filler words.
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
