Parag Agrawal was asked the obvious question about his company: web-scale crawling and indexing has been the preserve of Google and Bing, so why would a young company think it could compete? His answer began with what Parallel did not launch.
"We did not launch a search product first. We launched a search agent product first," he said on the Training Data podcast. That agent could go and crawl the web after a request arrived. Someone doing deep research, Agrawal said, has patience measured in a minute, and some products spend ten minutes on a task. That is a great deal of time in which to fetch pages, provided you have enough of a map of the web to know what to crawl first.
Agrawal, who ran Twitter before its sale to Elon Musk, founded Parallel Web Systems on the bet that agents would use the web a thousand times more than people ever have, and that the technology and business models behind search would have to be rebuilt for them.
What a search engine actually does
Agrawal laid out the mechanics for the hosts. A search engine crawls the web, finding URLs and reading them; it organizes what it reads into an index, so that when a query arrives it does not have to go and load pages; then it interprets the query, retrieves candidates and ranks them through several stages until hundreds of billions of pages have been narrowed to the handful a person wants. He calls it a billion-to-billion matching problem: enormous numbers of pages on one side, enormous numbers of queries on the other.
Google's own description of the process divides it the same way—crawling, indexing, serving—and adds a detail that matters here: an index is a selected, organized representation of content, including choices about which duplicate page to treat as canonical, not a complete copy of everything reachable.
Two things made that expensive machine hard to challenge, in Agrawal's telling. One was infrastructure cost. The other was feedback: knowing whether one result is better than another has depended on human ratings and click data that a newcomer cannot easily collect.
Parallel's position is that the second problem inverted. "Our view at Parallel is that human click data is a bug," Agrawal said, "and agent doing work with search should rely on agent feedback, not human feedback." Ratings can now be produced by experts far more cheaply, he argues, and the research that went into building models can be applied to indexing and ranking. Agents, in his framing, arrived as both a technology and a distinct customer.
The index as a latency optimization
The cost problem was handled by sequencing. An index, Agrawal said, "you can think of it as a latency optimization": it exists so you do not have to fetch pages once the query arrives. Give up that dimension and you are no longer competing with a search engine on day zero. You are competing with the alternative his early customers were actually using—outsourcing web research to people, who are, as he put it, far easier to compete with than a search engine on day zero.
By selling real work done on top of web data, Parallel could build its index incrementally as it served more customers.
The early work was unglamorous. Agrawal described search agents built for insurance underwriting and claims processing, sales teams enriching their data, and finance teams who had historically sent a data-collection job overnight to a set of people who would curate figures for a model the next morning. Parallel went after work that was already outsourced, which gave it something else it needed: empirical evaluations, drawn from customers' tasks, of what search for agents should look like—rather than tests invented in advance.
One host summarized the trade: turning off the crawl and spending compute at inference time instead. Agrawal agreed.
What the agent asks for, and what comes back
The interface changes when the customer is a model. People type two incomplete words with a typo and hope, or lean on autocomplete. "We're fundamentally lazy," Agrawal said. "Turns out, like, models less so." Longer, better-specified requests leave the engine less to guess. Parallel's Search API accepts a natural-language objective alongside queries—a way of saying what information matters—and returns source URLs, titles, publication dates where available, and selected excerpts.
Asked what happens when a developer sends a request, Agrawal described a pipeline. Models first interpret and enrich the request. Parallel maintains several indexes organizing different subsets of the web in different ways—a large one, a fresh one, others resembling a knowledge graph or a structured index—so the first layer decides which to consult and rewrites the query for each. Retrieval narrows tens or hundreds of billions of documents to thousands, then successive ranking layers, with larger models and more features at each stage, narrow those to specific paragraphs. Every query, he said, amounts to: "give me a thousand tokens from a trillion web pages on the web and make sure they're the right thousand tokens." Different versions of the API throw different amounts of compute at different points in that journey to meet different cost and latency limits.
Why the excerpt matters rather than the page comes through in an example about company financials. Ask for a public company's headline revenue and there is an authoritative answer in an SEC filing, perhaps on page 73 of a PDF that takes three and a half seconds to load. A summary page loads in a second and puts the number above the fold, and is right nearly all the time. People choose the second one, so the web fills with pages built for that choice—content the conversation labeled "pre-AI human slop," or, as Agrawal put it, "catering to a lazy human and being successful at SEO." He argues agents escape the trade-off, because the engine can lift the paragraph from the authoritative source into the agent's context window instead of making it fumble through a PDF.
His claim about the payoff is a comparison, and it is Parallel's: an agent using Parallel Search rather than calling Google at every step of a deep-research loop will, for the most part, use under half the tokens, and be more accurate and faster end to end. A context-limited agent that spends fewer tokens per search can do more work, or the same work more cheaply.
Latency last
Parallel organized itself around three dimensions—quality, cost and latency—and deliberately ignored one. "For the first couple of years, we said, let's focus, let's ignore latency, and let's just nail the other two," Agrawal said, on the reasoning that optimizing systems and distilling down to smaller models is "much more known art than unknown research."
Then it went after speed. Where the system had previously given itself three seconds to throw compute at a request, Turbo, announced on 13 July 2026, is reported by Parallel at a 200-millisecond median latency for a search request, priced at $1 per 1,000 requests, returning excerpts ready for a model. That figure covers the search request, not an entire research task; a single agent turn can fire off five to twenty searches, and heavier settings hundreds or thousands. Agrawal called Turbo "the fastest, highest-quality agentic web search on the market by a lot."
He also rejects the idea that model companies will absorb this work as a by-product of collecting pre-training data. Crawling for pre-training, he argues, is impatient by design: it is not worth waiting for slow JavaScript to render when there are easier tokens elsewhere. Parallel likes to crawl what others skip. "For us it is, because we're completionist."
That leaves the question of who builds and who buys. On the day of the recording, Agrawal said, Parallel announced work with Google Cloud as a search and grounding provider for its enterprise agent platform—an integration Parallel dated 16 July 2026, with access through the platform API, selection as a grounding source in Studio, and billing through Google Cloud Marketplace. Grounding means supplying current web information to a model as it answers. Attach web search to an agent built on Google Cloud, Agrawal said, and "your options are Google Search or Parallel Search." He expects several such partnerships, each looking somewhat different.
His account is a concrete infrastructure case of an argument developed elsewhere in our coverage, in the idea page The inner loop: why AI advantage may sit outside the model—that durable advantage comes less from the model than from an organization's own cycle of deploying, measuring and improving. Parallel's version of that loop was to sell agents to customers whose work it could learn from, and let the index catch up.
Agrawal is not claiming the shift has happened. Asked whether agents now issue more queries than people, he said no, not yet: "we are very, very early in agent adoption."