Parallel’s Parag Agrawal: Building a New Web for AI Agents

Parag Agrawal explains that Parallel Web Systems is building technology for agents to search and use the web, based on the expectation that agents will perform searches roughly 1,000 times more often than humans. Humans use browsers, search engines, clicks, and browsing; agents need equivalent access but at a radically larger scale. Parallel therefore aims to redesign both search infrastructure and its business model. Agrawal contrasts this with his Twitter experience: Twitter was a mature, post-product-market-fit company with hundreds of millions of users generating feedback, while Parallel is building for a customer—the agent—that is only now emerging.

Agrawal describes ordinary web search as finding information whose location the user does not know. A search engine crawls URLs, reads and organizes pages into an index, then retrieves and ranks a tiny number of relevant results from hundreds of billions of pages. The system must match hundreds of billions of possible queries against hundreds of billions of pages, a “billion to billion” problem. The hosts note that Google and Bing historically had the resources and feedback to operate at this scale. Agrawal argues that models and agents now make a new entrant more viable.

Parallel initially avoided competing with Google on day one. Its first product was a search agent that crawled the web after a query arrived, sometimes taking a minute and, for deeper research, up to 10 minutes. That latency was acceptable when the alternative was outsourcing research to humans. Early customers used it for insurance underwriting and claims processing, sales-data enrichment, and finance workflows that had previously sent requests overnight to human researchers. Those uses generated real examples of what agents needed, while revenue funded an increasingly broad index. Agrawal calls this trading crawl completeness for inference-time computation.

Agrawal says Parallel views human click data as “a bug” for agent search. Instead of judging results by what humans click, agents can provide ratings, and models can compress information and help with indexing and ranking. This makes the problem more tractable because expert-quality ratings can be created more cheaply. Parallel also focuses on pages and data that model companies may not collect for pretraining: slow JavaScript-heavy pages, fresh information, and material that is expensive to crawl relative to the tokens obtained. Its index is built incrementally as customers create demand.

Agrawal rejects the “neolab” label because Parallel’s output is not a foundation model; it is a complement that gives models and agents better information and new capabilities. The system tries to reduce trillions of web tokens to the roughly 1,000 highest-signal tokens relevant to a query. Until recently, Parallel allowed about three seconds for this work; a newly demonstrated product called Turbo performs it in 20 milliseconds. The research focuses on memory hierarchy, adaptive ranking architectures, training data, and allocating limited compute to maximize quality under cost and latency constraints.

Humans often submit short, incomplete, typo-filled keyword queries because they are lazy and rely on autocomplete. Agents can provide longer, better-specified requests, giving search systems less intent to guess. Agrawal also explains why pages that summarize authoritative documents often rank well: a company filing may contain the correct revenue number on page 73, while an SEO page puts that number above the fold and loads quickly. Parallel can instead extract a precise passage from the authoritative source and place it directly in an agent’s context, avoiding the human tradeoff between authority and convenience.

When a developer sends a query, a model enriches and rewrites it for several specialized indexes, such as a broad index, a fresh index, or a structured knowledge index. Retrieval narrows billions of URLs to thousands of candidates; successive ranking stages and larger models then select the most useful excerpts and ultimately about 1,000 tokens. Different API versions spend different amounts of compute to meet latency and cost limits. Agrawal says search is never simply “finished”: the question is whether an incremental quality improvement is worth its cost, especially as cheaper intelligence enables better matching.

A normal AI answer may already trigger five to 20 searches, while more intensive systems can trigger hundreds or thousands. Agrawal gives the example of monitoring a portfolio of 10,000 small-business borrowers: an agent could replace a monthly human review with a weekly dashboard and action list, producing hundreds of thousands or millions of searches. A meeting-preparation agent can combine internal company information with Parallel’s web data and perform tens or hundreds of searches for every meeting. He expects background agents to run whenever their incremental value exceeds their compute cost, though current users are still outliers.

Agrawal says advertising made free search, social media, and web content possible because it supports differential pricing: most queries lose money while valuable commercial queries subsidize them. That model assumes human attention and identifiable visits. Agents break the loop because publishers may see anonymous automated access without knowing whether it produces a subscription or valuable work. Fixed licensing deals with model companies also do not scale if inference grows sevenfold year after year. Parallel proposes using Shapley values to estimate how much each source improves an agent’s result, paying more for unique content and for content used in high-value work.

Agrawal describes three stages: agents first use web search as a familiar tool; more advanced systems then coordinate multiple specialized agents; eventually the web shifts from pull to push. Instead of asking an agent to find something at a point in time, a user could ask Parallel to monitor the web and call the agent when a relevant change occurs. Triggers might include satellite imagery, customer commentary, another agent finishing work, or a new human insight. Agrawal says that if useful work can be done today and GPUs are available, agents should do it now; tomorrow’s work will increasingly be triggered by changes detected across the web.