Parag Agrawal on Web Search and Business Models for AI Agents

Parag Agrawal describes Parallel as a search provider for AI agents. Agents working for people or organizations need information from the web, he says, but their requirements differ enough from those of human searchers to call for new technology and sustainable business models.

His founding hypothesis is that agents could eventually use the web a thousand times more than people. Agrawal treats that figure as a way to reason about scale, not a measured current usage rate: if demand grows by orders of magnitude, both the underlying system and its economics would have to change.

In that scenario, he estimates, search would need to become roughly ten to a hundred times more compute-efficient. Otherwise multiplying today’s search cost by a thousand would be prohibitive. The exact scale and efficiency improvement remain his projections.

Human search typically involves a few short keywords, a wait of around half a second or a second, and a list of links that the person explores. Agrawal expects an agent to supply a more explicit description of its task and consume structured information or files instead of blue links.

The time budget also varies. A voice agent may need a result in roughly 100 milliseconds while a user waits; a background research agent can take much longer for a better answer. Search must accommodate those different constraints rather than optimize only for the familiar human interface.

The model using the result matters too. A small, inexpensive model and a larger, more costly model impose different trade-offs: the system must decide how much noise it can pass through and how much work to do before returning information.

Agrawal describes a retrieval funnel. Starting from a vast number of web documents, a system spends almost no compute on most, a little on a smaller candidate set, and progressively more ranking effort as it narrows the results to information that fits into the model’s context.

That process spends search compute to save model compute. With a cheap model, it may be acceptable to pass along more material; for an expensive model, more search and filtering may be worthwhile if they prevent wasted model time or repeated searches.

Asked how Parallel handles different preferences for accuracy, latency and cost, Agrawal says an application developer or agent can configure the API. The caller can identify the model and choose among modes intended for different workloads; supplying that context may improve results, he says.

He describes a fast mode intended for voice agents and a more compute-intensive mode for expensive background tasks. A longer search can sometimes lower the total time and cost if it helps an agent answer with fewer subsequent searches.

Parallel’s common use cases are knowledge work, Agrawal says, including coding, legal research, insurance underwriting, sales and science. He estimates that coding invokes web search for only a minority of prompts because much coding work depends on internal files and a repository.

Legal work, by contrast, can depend on finding case law and checking facts about people or companies. Agrawal says some tasks require comprehensive searches to establish confidence that relevant information has not been missed, even when private data also matters.

He welcomes more useful personal assistants and other agents because Parallel’s business depends on agents consuming web information. In his view, better and cheaper models enable more agent use cases, which can increase the opportunity for search.

Stebbings asks whether smarter models will search less. Agrawal distinguishes reasoning from factual memory: a model may remember common facts but cannot be expected to retain every detail in its training data, which he characterizes as a form of lossy compression.

Some facts are absent from training altogether, and information can become stale. Agrawal also expects smaller models optimized to preserve useful performance to retain less parametric knowledge, leaving a role for external retrieval.

He predicts two trends at once: frontier models may grow larger, while smaller models reach a fixed level of performance more efficiently over time. He expects high-value problems to justify spending more compute on better answers even as routine capabilities become cheaper.

Agrawal says he does not know how open and proprietary models will divide future tokens and spending. He questions a precise 90% split in either direction because model costs, application needs and the development of open models remain uncertain.

He hopes strong US-built open models and competition between their developers will emerge, but says the outcome is unclear. His view is conditional on how the technology and market develop rather than a firm forecast of any particular provider’s success.

On model routing, he separates choosing a model from choosing a GPU and vendor serving it. He says routers have real value while fast-growing customers struggle with uneven capacity: they can switch providers, and sometimes models, to meet demand and service requirements.

Whether the routing layer remains as valuable depends partly on future supply and demand for GPUs and tokens, Agrawal says. He then identifies data and unique insights as areas that could become much more valuable over the next few years.

His premise is that cheap intelligence becomes more useful when combined with private knowledge, another party’s data and public information. The challenge is to create transactions that let the parties combine those inputs and pay for their contributions.

Agrawal emphasizes inference-time access, not only licensing data for model training. A venture investor might combine personal notes and experience with a paid information service and public data; an agent acting for that investor needs a lawful, practical way to use the same valuable information.

A per-seat subscription designed for a human can be awkward when many agents need access. Agrawal argues for compensation based on agent use or value, so data owners have an incentive to provide access instead of repeatedly blocking automated clients.

Stebbings asks whether companies should let agents into their services, pointing to different approaches in e-commerce and travel. Agrawal expects businesses eventually to accommodate agents, but says the terms and strategic choices will depend on their incentives and market power.

He does not declare a company right or wrong for restricting an agent. A platform might retain customers through its own assistant or negotiate access on different terms; it could also lose transactions if customers move to other agents and it fails to offer a useful alternative.

Agrawal is uncertain what share of shopping people will delegate. Some consumers enjoy choosing products themselves even if they hand other chores to an agent. Stebbings is much more convinced, based partly on his own intensive use of a personal assistant.

The change raises a problem for advertising, Agrawal says. A publisher with an ad-supported page earns money when a person sees an ad, but an agent extracting the information may not expose the advertisement to anyone. Current ad formats therefore do not directly translate to agent consumption.

He describes Parallel’s intended response as an ‘AdSense for agents’: paying content owners a variable amount when agents derive value from their information. His aim is to align incentives so publishers have a reason to make useful content available.

Stebbings asks why a publisher would cooperate if other agents still consume its work without paying. Agrawal says a content owner could provide a feed to a party offering acceptable terms while using available technical or legal measures to restrict others.

Would publishers end up with a smaller core business, as some creators experienced during the shift to streaming? Agrawal says he does not know, but expects substantially greater agent usage to create a larger potential market. He also acknowledges there would be winners, losers and adjustments.

In his view, some content could become more valuable and some more commoditized. Whether the overall market grows as he expects depends on agents actually producing useful work and on business models that return value to the people supplying information.

Asked about margins, Agrawal says Parallel is an infrastructure company focused on quality, cost and latency. He claims its technology can achieve comparable quality with a fraction of the compute used by systems built for human search; the interview does not independently verify that comparison.

He says future margins will depend substantially on competition and market structure, which remain unsettled. He does not consider paying content owners inherently fatal to margins because useful information can reduce the compute needed to achieve a desired result.

His conceptual example is an agent with a fixed budget: if removing a publisher’s information reduces answer quality, the content has made a marginal contribution. Agrawal says Parallel tries to estimate that contribution and share a meaningful part of its value with the owner.

For market sizing, he proposes looking at inference spending on agents and excluding training and media generation. He estimates that web search could account for roughly 5% to 20% of the relevant GPU spending. This is his working assumption, not an observed market-wide ratio.

If inference demand grows and Parallel maintains its share, he expects its revenue opportunity to grow alongside it; winning share would change that outcome. Stebbings and Agrawal work through hypothetical revenue and valuation figures, explicitly relying on assumptions about market growth and share.

Agrawal says a very large valuation within a few years is possible, but not guaranteed. He identifies a broad risk that agents fail to deliver enough value after society has invested heavily, as well as company-specific risks around execution.

Even if agent adoption expands, Parallel would need strong technology, agreements with content providers and the trust of customers with substantial inference spending. Agrawal notes that the future distribution of customers and model providers is itself uncertain.

Stebbings raises commoditization, noting that several providers can be close on public benchmarks. Agrawal responds that lower prices are desirable if search is to scale dramatically, and argues that current prices are poorly matched to some low-cost agent workloads.

His example is an inexpensive model running repeated deep research. If search costs dwarf the model cost, he says, the allocation is backwards; in his view search prices may need to fall by an order of magnitude or more for such use cases.

For a more expensive model, Agrawal is willing to spend more on search if it improves the answer and reduces later work. He expects providers to offer both very cheap search for inexpensive models and more intensive search for high-value tasks.

He connects historical search pricing to an advertising-supported human market, in which a search provider could justify higher infrastructure cost against ad revenue. Agrawal claims Parallel can offer similar quality much more cheaply and expects further reductions, though these are company claims and forecasts.

Public benchmarks do not tell the whole story, he argues: some can become saturated or be memorized, and a quality score does not show the cost of producing the result. He says buyers should consider both usefulness and price rather than treating a close leaderboard as proof all services are interchangeable.

Stebbings describes using an agent to watch European company registries for young founders. Agrawal uses that example to explain why persistent monitoring might be more efficient than an assistant repeatedly waking up to run the same searches.

Parallel already crawls the web, Agrawal says. If its system knows a user’s long-lived monitoring request, it can evaluate new changes as they arrive and spend more compute only when a change looks relevant, rather than repeatedly searching an unchanged world.

He describes the Monitor API as an event-triggered alternative to periodic polling. A user supplies the conditions for a notification; the system screens changes cheaply and can perform more expensive checks on promising events before alerting the agent.

This can reduce repeated work when relevant events are rare, though results still depend on crawling, matching and false positives. Agrawal sees a broader shift from pulling information on a schedule to pushing relevant web events to persistent agents.

In his hypothetical future, an agent completes work it can do now and waits for something new: an email, another agent’s result, a change on the web or a person’s new idea. Those events give it a reason to resume. This is his vision, not a description of universal adoption today.

Stebbings asks how to stop a goal-directed agent from pursuing an instruction through harmful means, using an unauthorized booking change as a hypothetical. Agrawal says the risks are complex and distinguishes experiments during reinforcement learning from the behavior of models released to users.

He says some previously reported incidents involved models during training rather than fully aligned released systems, while noting he does not know the details of a newly mentioned incident. His interpretation is that alignment work can reduce certain risks, even though alignment remains unsolved.

Agrawal argues that the design of training environments can reduce opportunities for harmful behavior during reinforcement learning. After deployment, he sees both accidental misuse and deliberate malicious use as concerns, and says model alignment is not guaranteed to withstand an adversary.

The discussion turns to cybersecurity. Stebbings warns about malicious actors using capable agents; Agrawal says systems need protection and that model builders have a responsibility to minimize foreseeable harm from what they create.

He acknowledges that stochastic systems cannot be guaranteed perfectly safe. But he objects to treating every security incident only as proof of model power: an incident may also reveal safeguards that could have been stronger. He welcomes detailed retrospectives that make those failures visible.

Agrawal’s broader fear is that AI could be useful overall yet spread unevenly, with benefits too concentrated and the transition painful for many people. He worries that society may adapt more slowly than the technology changes.

The speakers discuss giving agents authority to spend money or use accounts. Agrawal says they inhabit an unusually early-adopting bubble: technical tools for delegated payments may arrive before most people are socially comfortable giving agents access to financial accounts or email.

Asked about wealth concentration, he says some disparity is compatible with his worldview but worries the trend could go too far. He is unsure where the threshold lies and hopes social forces can address the problem without extreme disruption.

Stebbings asks whether vertically integrated companies will capture most AI value. Agrawal sees benefits in integration but warns that a company which refuses to work with others may limit its reach in a rapidly changing market.

Parallel itself integrates parts of the search stack, but stops at an API so it can serve many different agents, Agrawal says. Its bet is that a common information-retrieval problem exists underneath applications in many domains, including those built by vertically integrated companies.

Asked what founders might learn from Elon Musk, Agrawal first says he has many disagreements with him. He praises Musk’s sense of urgency and ability to compress timelines, arguing that ambitious expectations can help people do more than they thought possible when paired with inspiration.

On the current Twitter product, he singles out the project formerly called Birdwatch, now Community Notes, as an idea he liked and is glad people continue to develop. He does not offer a broad endorsement of the platform’s direction.

Agrawal says he generally does not make venture investments: his wife works in venture capital, and their compliance process makes investing cumbersome. He also sees no reason to assume his brief exposure to a founder would make him a better investor.

From Vinod Khosla, he says he learned the value of technical intuition and a long-term technical aspiration. After solving one problem, the discipline is to look ahead and make informed bets on the next several technical challenges.

Agrawal has also changed his view of sales and marketing. He began with a strongly product-and-technology-focused mindset; running Parallel has shown him week by week how much capable sales and marketing can contribute.

Asked about competing products, he distinguishes Perplexity’s more vertically integrated consumer offering from Parallel’s search infrastructure business. He sees Exa as more directly in the same search-infrastructure market. These are his characterizations of the companies’ positions.

He recalls a venture meeting with Josh and Todd, who flew out to spend time with him after deciding to invest. Looking ahead ten years, he says he is excited by change and by the possibility that a company can influence what happens to the open web and content owners.

His final prediction is that persistent agents will run on people’s behalf much more routinely than many currently expect. Agrawal acknowledges that even among technology-focused peers the idea is not universally accepted; it remains his expectation rather than an established outcome.