AI Agents Are Killing Human Attention; Search Must Become Programmatic
By Nikhil Raghavan ·
Web search is pivoting from capturing human clicks to executing complex workflows driven by machine computation.
The idea that agents will perform tasks using the web 1000x more frequently than humans fundamentally changes how we must think about the internet. This was the central premise of the recent "Sequoia Capital" podcast episode featuring Parallel’s Parag Agrawal: a shift from human-centric search to agentic computation.
On the podcast, Agrawal and his co-panelists detailed an architectural pivot away from relying on human click data, arguing that future web search must be driven by machine feedback. The core technical challenge, they asserted, is the "billion to billion matching problem." Parallel’s strategy, as presented, involves building a complementary search index—a powerful adjacency for LLMs—rather than attempting to become a full model company itself. They achieved massive optimization gains, reducing search latency from three seconds down to 200 milliseconds by organizing information across memory hierarchies. Furthermore, they outlined fundamental changes in business models: the traditional ad-based revenue stream, built on the scarcity of human attention, is eroding.
The End of Human Scarcity and Attention Economics
The most disruptive claim presented was that the foundational economic assumption underpinning the internet—that human attention is scarce—is now obsolete. As agents become primary users of web data, the value proposition shifts from capturing clicks to efficiently executing complex workflows (like insurance underwriting or claims processing). The historical pattern here is cyclical: every major technological leap—from dial-up to mobile, from desktop search to voice assistants—has destroyed the previous revenue model and created a new one. What we are seeing now is not an iteration of search; it is a systemic replacement of human intent with programmatic capability.
Defeating the Static Content Model
The strongest objection to this agent-first vision is that content creators, particularly those who rely on established SEO practices, can simply adapt by creating highly structured "dual published" content—designing material specifically for both human and AI extraction. They could argue that traditional publishing formats, optimized with clear headings and data points "above the fold," are sufficient to satisfy agent needs without requiring a dedicated index buildout like Parallel’s.
However, this argument fundamentally misunderstands the scope of the problem. The proposed solution assumes agents will simply read structured data; it fails to account for the need for specialized, multi-layered retrieval that goes beyond simple extraction. Agrawal himself noted that agentic search involves enriching queries and crafting specific requests across multiple indexes (fresh, knowledge graph, structured). This deep functional adjacency—the ability to combine several specialized information sources into a single answer—is not achievable by simply structuring an article; it requires the infrastructure Parallel is building. The gap between merely presenting data and programmatically connecting disparate datasets remains massive.
Infrastructure as the New Moat
The discussion on payment mechanisms cemented the necessary shift in value capture. Traditional fixed-fee contracts for training data are failing because AI inference growth (projected at 7x this year, 7x next) far outpaces linear contract increases. The theoretical framework of Shapley values—determining how much incremental value any single piece of content adds to a final output—provides the necessary conceptual basis for an "auto bit" system. This moves payment away from fixed access rights and toward dynamic ROI measurement.
The web is moving towards Level 3: the "push" model, where information proactively signals availability ("Call me if this happens..."). The central policy challenge here is regulatory: how do we govern a network that operates on continuous, automated signal exchange rather than discrete human requests? If the core value accrues to those who build and maintain the complementary infrastructure—the specialized index and matching algorithms—then these entities are not merely service providers; they are becoming essential utility layers.
The future of information retrieval is not about better search queries or smarter ad placement; it is about building a universal, incentive-aligned computational plumbing that can reliably connect diverse data sources in real time. The winner will be the entity that solves this complex optimization problem around quality and latency, making the physical infrastructure itself the most valuable commodity on the internet.