> ## Content Index
> Fetch the complete content index at: https://www.signalful.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Agents Will Use the Web 1,000x More
- URL: https://www.signalful.com/agents-will-use-the-web-1-000x-more/
- Published: 2026-10-02T13:49:00.000Z
- Updated: 2026-10-09T19:40:19.000Z
- Description: Parag Agrawal built Parallel on one bet, that AI agents will use the web 1,000 times more than people do.
- Author: Signalful Editorial
- Tags: Builders, Parallel, Parag Agrawal, AI Agents, Search, Business Model, #sf-0003, #yt-wTxb_whJR00, #Import 2026-10-06 15:50

*Parag Agrawal, founder of Parallel and former Twitter CEO, interviewed by Harry Stebbings (20VC)*

**Summary:** Parag Agrawal started Parallel on one written line: agents will use the web 1,000x more than humans, so the web needs new technology and new business models. At that scale, search built for people is too slow, too expensive, and priced off Google's ad economics. Ads also stop paying publishers once agents do the reading. Agrawal's answer is web search priced for agent workloads, with publishers paid for what their content adds to an agent's result.

1. **The 1,000x Premise.** Agrawal says "no tech built for a certain scale survives three orders of magnitude." That one assumption shaped how Parallel builds everything underneath. If agents search 1,000x more and each search costs what it does today, the compute bill is absurd. Per-search compute has to fall 10x to 100x before agent-scale search makes economic sense.
2. **Agent-Shaped Queries.** Agents ask for search in a different shape from humans. Humans type 3 short keywords, wait half a second to a second, and click through 10 blue links. An agent sends a full sentence describing the task. A voice agent needs an answer in 100 milliseconds, a background agent will wait for the best answer, and either one wants tokens or files back instead of links.
3. **Compute Allocation.** Web search spends compute to save model compute. The job is to shrink a few trillion documents to about 1,000 tokens for the model's context window, using bigger rankers at each stage. A cheap model can tolerate a little more noise in context, while an expensive model should get heavily filtered results. Parallel sells this as modes: Turbo for voice agents, Advanced for costly background agents that think for 20 seconds, and an API parameter that lets callers say which model is asking.
4. **Facts Outside the Model.** Smarter models won't shrink search demand, because parametric memory is lossy compression. A model knows who was president in a given year; it can't tell you when a Parallel employee graduated, even if that fact was in its training data. Coding calls web search in only about 5% of prompts because the context is the codebase. Law, insurance underwriting, sales, and science lean hard on search because they need case law and facts about companies and people.
5. **Ads Break With Agents.** "Ads don't work with agents in their current form," Agrawal says. An ad-supported page makes money when people show up and see ads; when agents show up, nobody sees ads and the publisher earns nothing. Left alone, every publisher blocks agents. Parallel's fix is "an AdSense for agents" that pays content owners a variable amount each time an agent benefits from their page.
6. **Paying Marginal Contribution.** Parallel pays publishers based on how much their content improves an agent's answer. Say an agent spends $1 on a task, and removing one publisher from the index drops quality to what a 90-cent run would produce. That publisher contributed 10 cents, and Parallel pays out "a decent chunk" of it. For an equally good product, Agrawal would rather pay content owners than spend the same money on inference.
7. **Data Priced at Inference Time.** Agrawal thinks data is underpriced now and will be worth far more in 3 to 5 years. His example is PitchBook, which charges per seat, while a VC's agents either can't reach the data or scrape it through a browser, possibly against terms of service. If nobody works out how to pay for data when agents use it, data owners and agents stay locked in a "cat and mouse game." Agrawal expects every company to let agents in eventually; the fight is over terms.
8. **Mispriced Web Search.** Web search sells for roughly $10 per 1,000 searches, a price set when Google's ad CPMs made search costs irrelevant. A cheap model running deep research at today's prices spends 80% to 90% of its budget on search and 10% on the model, which Agrawal calls "entirely silly." Parallel charges $1 per 1,000 at comparable quality, against 7 to 14 dollars elsewhere, and he sees another 10x drop ahead. He welcomes a race to the bottom on price and gives little weight to public benchmarks like BrowseComp, which he says half the models have memorized.
9. **A Slice of Inference.** Agrawal estimates 5% to 20% of the GPU spend on running agents will go to web search. If inference revenue reaches $200 billion to $300 billion, that's a $10 billion to $60 billion market. A 33% share would mean about $6 billion to $7 billion in revenue, which he says supports a $100 billion company. The risks he names: agents fail to deliver and society overbuilds GPUs, Parallel loses on technology, it can't sign enough content partners, or it doesn't win the customers who end up running most inference.
10. **Search From Pull to Push.** Always-on agents make polling the web wasteful. An agent that checks every 6 hours for, say, new European companies with an under-25 founder burns compute on runs that mostly find nothing. Parallel's Monitor API, which Agrawal calls Google Alerts made smart, flips this: Parallel already crawls the web constantly, so it spends compute only when something changes. He claims that does the same job with 100x to 1,000x less compute than polling.
11. **Hacks as Embarrassments.** Agrawal says AI labs should treat publicized agent hacks as embarrassments. Each incident shows the model is powerful, and also that "we did not guardrail them enough." Most reported incidents came from models still in RL training, which he considers fixable by building safer training environments; the bigger worry is that a model's alignment "is not adversary proof" against people who want to cause harm. He worries even more that AI proves net positive and society still diffuses it badly, stays too concentrated, and makes the next few years "really really rough."
12. **Unreasonable Expectations.** Agrawal has "lots of disagreements" with Elon Musk but admires Musk's urgency and ability to compress time. Most people sandbag themselves and set lower expectations than they can meet, so being inspired and pushed at once gets more out of them. Agrawal's own change of mind in the last year points the same way: he started believing only technology and product mattered. Now he sees the week-on-week difference that strong sales and marketing make.

Watch the full video at [https://www.youtube.com/watch?v=wTxb\_whJR00](https://www.youtube.com/watch?v=wTxb%5FwhJR00&ref=signalful.com). Read the full transcript at [https://www.usetranscribe.io/yt/wTxb\_whJR00/ads-business-model](https://www.usetranscribe.io/yt/wTxb%5FwhJR00/ads-business-model?ref=signalful.com).