Dhruv Batra argues that the less standardized parts of the web are unlikely to acquire useful APIs quickly enough to serve every agent task. The talk contrasts straightforward flight-search integrations with restaurant menus and school procurement information scattered across pages, scanned PDFs and image galleries. The point is not that APIs are undesirable: Dhruv Batra explicitly favors them when an appropriate service already exists.
Dhruv Batra explains why reading a page’s initial HTML can miss information that a person sees in the browser. Content may arrive through later network requests, while availability or other state can be expressed through rendering logic, disabled controls and visual styling. These examples motivate agents that can interpret the rendered interface rather than depend exclusively on static text extraction.
Dhruv Batra proposes combining visual interaction with code where each is useful. An agent may click or scroll when needed, use scripts for suitable operations and inspect the resulting interface to check what happened. The discussion also considers parallel browser sessions as a way to distribute independent tasks, while acknowledging that accuracy, latency and cost remain practical constraints.
Dhruv Batra predicts that an additional agent layer could eventually expose structured task results over existing websites without requiring each institution to rebuild its systems. This is a proposed direction for web automation, not an established universal capability. The talk’s product performance and future cost claims are not independently verified here.
Watch on YouTube
