20VC: Five Predictions for a World of Agents | The Ads Business Model Will Die | Biggest Lessons from Working with Elon Musk at Twitter with Parag Agrawal, Parallel

Summary

Parag Agrawal makes the case that the internet is shifting from human search to agentic search, where systems will query the web far more often, with different inputs, outputs, and latency constraints than people. He says that means search must become much more efficient, with specialized APIs for voice agents, background agents, and monitoring workflows. A major theme is that current ads-based content monetization breaks when agents consume content without seeing ads, so Parallel is effectively building a new payout model for publishers and data providers. Agrawal also discusses model scaling, arguing frontier models will keep getting larger while smaller models will hit fixed performance targets, and that routing, data access, and inference economics will matter more as agents proliferate. Later in the conversation he shifts to safety and platform design, saying alignment is still unsolved, builders should take responsibility for harm reduction, and the biggest societal risk may be poor diffusion and concentration of AI benefits rather than raw capability.

Notes

• Parallel is positioned as search infrastructure for AI agents, not a consumer search app.
• The founding thesis is that agents will search the web 1,000x more than humans.
• If that usage pattern holds, the web search stack must become 10x to 100x more efficient to be economically viable.
• Agent search has different inputs, different time budgets, and different outputs than human search.
• For voice agents, latency budgets can be around 100 ms rather than the 500 ms humans tolerate.
• Parallel’s API can be parameterized by the calling model, letting the stack adapt to the agent or application.
• The product has multiple modes, including a very fast low-cost path for voice and a slower advanced mode for expensive background agents.
• Coding is a large inference category, but only about 5% of coding prompts trigger web search.
• Web-search-heavy professions include law, insurance underwriting, sales, and AI science.
• He expects frontier models to keep getting bigger over time.
• Smaller models, in his view, will keep closing the gap to any fixed performance target every six months.
• He says there is no visible end to scaling laws yet, and more compute can still improve results.
• Routing still has value because customers need flexibility across model choice and GPU vendor/capacity.
• Data is one of the most undervalued future assets, especially data useful at inference time.
• He thinks the market still lacks a good mechanism to pay for unique insight or proprietary data.
• He cites PitchBook as an example of a valuable dataset that should be accessible by agents and compensated by usage.
• He argues every company will eventually need to let agents in.
• The customer is shifting from human to agent, which changes product design and access policy.
• Agrawal says ads do not work with agents in their current form.
• Parallel is effectively building a replacement to the ads business model for agent consumption of content.
• He frames the market as a growing pie, not just a redistribution of existing web value.
• Parallel claims it can deliver the same quality answer with far less compute than a human-search stack.
• He estimates 5% to 20% of inference GPU spend for agent workloads may need to flow into web search.
• That framing makes web search a large adjacency to the inference market, with major revenue potential if agents scale.
• He sees failure risk in whether agents deliver and whether Parallel captures enough share.
• He wants a race to the bottom on web search pricing to support 1,000x usage.
• If agentic workflows consume a lot of web search, he expects web search to become a much larger share of spend than the model itself.
• He says public benchmarks are saturated and can be misleading because models may have memorized them.
• Parallel’s search is priced at about $1 per 1,000 searches for that quality level, versus roughly $7 to $14 for many competitors.
• He argues the web is moving from pull to push via event streams.
• Parallel’s Monitor API is meant to function like Google Alerts for LLMs.
• The monitoring product can save compute by polling less often and escalating only when needed.
• He says alignment is still unsolved, even though post-alignment systems appear safer than partially trained ones.
• He places responsibility on model builders to minimize misuse and harm.
• The larger worry, he says, is not just model capability but AI benefits concentrating too much.
• Vertical integration has merits, but being too rigid can box a company out.
• He admires Elon Musk’s urgency and ability to compress time.
• His lesson from Twitter is that technical excellence alone is not enough; sales and marketing matter week to week.
• He distinguishes Exa as a direct search competitor, but not Perplexity, which he sees as a vertically integrated product.
• He sees Parallel’s role as tied to the changing open web and the incentives of content owners over the next decade.
• He thinks people are not yet socially ready to hand agents their bank access, email, and passwords at scale.