Human Click Data Is a Bug: Rebuilding Search for a Web of Agents

On Sequoia Capital’s Training Data, Sonya Huang and Andrew Reed sit down with Parag Agrawal, the former Twitter CEO who sold the company to Elon Musk and then went back to founding. He now runs Parallel Web Systems, which he describes as technology that lets agents search and use the web the way humans have long used browsers and search engines. The company started with a bet that agents would do that work a thousand times more than people ever have. If the bet is right, both the retrieval stack and the business models sitting on top of it have to be rebuilt.

The multiplier is already visible in ordinary products. A typical search agent, even when it answers in a few seconds, still runs five to twenty queries underneath. In ChatGPT, a prompt at a middling reasoning setting already triggers five to ten searches; turn the setting up and the same turn can spawn hundreds or thousands. Three years ago Agrawal figures he ran twenty or thirty Google searches a day. Count the agents running under his name now, and he puts the load at a hundred to a thousand times that — more if you include company-wide jobs that are not assigned to any one person.

He is quick to say he is an outlier. Step outside the Silicon Valley bubble, he tells the hosts, and most people have not even heard the new vocabulary. The hour that follows is an argument for why a young company can still take on web-scale search, why human click logs are the wrong training signal for this customer, and what happens to an internet funded by ads when the visitor is no longer a person.

Reed, making his Training Data debut, asks Agrawal to step back from Silicon Valley’s surface-level picture of search. The job looks simple: you do not know where something is on the web, so you go to an engine and hope it surfaces the answer in the most convenient place. Behind that, the engine crawls every URL it can find, reads the pages, and organizes them into an index so a query does not start by loading pages from scratch. Then it understands the query and runs several stages of retrieval and ranking, taking hundreds of billions of pages down to five, ten, or ideally one result.

Agrawal’s compression of the whole problem is that it is a billion-to-billion matching exercise. Hundreds of billions of pages sit on one side, hundreds of billions of queries accumulate on the other, and the engine has to matchmake across the two. Huang’s reaction is the obvious one. Full-web crawls and indices have, for a long time, been a problem only Google and Bing actually finished. Why would a young company spend that kind of money?

It is expensive in the long term, he grants. Three years ago the old moats were feedback: human ratings and click data that a startup could not collect at scale. Large models changed the cost of expert ratings, and Parallel’s view is that agents should be scored on agent feedback anyway. The other half of the plan is not to buy the entire index on day one. If you can grow a more sophisticated index as you solve problems for more customers, the infrastructure bill rides the customer base instead of preceding it.

Building Parallel has also meant unlearning Twitter. In leadership roles there, he was running a post-product-market-fit machine: hundreds of millions of people, thirty-plus minutes a day, feedback loops from real usage. Parallel is the opposite, a pre-product-market-fit company built for a customer that will show up in a few years. The team is learning every week how to build for someone who is not here yet.

The first product was not a search engine. It was a search agent that crawled the web after the query arrived. Deep-research buyers will wait a minute, and some of Parallel’s products take ten; that is enough time to fetch a lot of pages if you have a map of what to prioritize. An index, in this framing, is mostly a latency optimization. Give up that dimension and live crawling covers the gaps. Huang boils it down to trading inference-time compute for crawl time.

The competitive insight sits in one comparison. Competing with a search engine on day zero is hard; competing with humans who sit in front of search engines and curate data is much easier. Early customers were exactly that outsourced work: insurance underwriting and claims, sales data enrichment, finance teams who used to ship a pack overnight to people who would clean it before a modeling exercise. Parallel took the jobs in order to collect evals from real work, so the team would learn what agent search should look like from empirical use cases rather than theoretical benchmarks.

The growth math is a multiplier. A lender with a portfolio of 10,000 small businesses used to run a monthly human process over web data to watch credit risk; an agent can run it weekly and fire hundreds of thousands or a million searches per pass. A developer writing one program is a 100,000x to 1,000,000x search multiplier. Agrawal uses Notion’s agent, wired to Parallel’s internal data plus the public web, to build meeting-prep docs. One prompt, then tens or hundreds of searches per meeting, and every new agent multiplies again.

The company is organized around three axes: quality, cost, and latency. For the first two years they ignored latency and tried to nail the other two, because distilling large models into smaller ones is more known art than unknown research. Once they had the best quality at every price point, they turned to speed. The Turbo product they had just shipped takes the “thousand tokens from a trillion pages” read-time compute from three seconds down to 200 milliseconds. Agrawal calls it the fastest, highest-quality agentic web search on the market, by a lot.

The line Agrawal keeps returning to is company doctrine. “Our view at Parallel is that human click data is a bug.” An agent doing work with search should rely on agent feedback, not human feedback. Human search is distorted by laziness: two incomplete words and a typo, then hope the engine guesses. Agents send longer, better-specified queries with fewer typos, so the engine has less to infer. Training an agent search stack on human clicks is using the wrong signal for the actual customer.

Reed asks whether Parallel, with a roster of AI researchers, is a “neo-lab.” Agrawal will not take the label. A neo-lab’s output is a model; Parallel’s output is a complement to a model, something that multiplies on top of it so the model, or the agent built with it, gets superpowers. When another lab ships a better model, that lab has a higher hill to climb. When someone ships a better model, Parallel unlocks more use cases. They are not pre-training enormous models. Their job is the opposite: compress the problem into tiny ranking models.

He restates the query the search engine actually receives: give me a thousand tokens from a trillion web pages, and make sure they are the right thousand. Until the week of the taping they gave themselves three seconds of read-time compute; Turbo now does it in 200 milliseconds. The research is how to organize information across the memory hierarchy, how to train a model that picks those tokens, and how to spend a limited compute and latency budget. They take a pretrained model, adapt its architecture, and squeeze the best output they can from that budget.

Huang’s industry view is that search is so fundamental that model companies will just own it, especially since pre-training already requires an expensive crawl. Agrawal does not see the empirics on the ground. A crawl built for pre-training is not that useful for a fresh index, because model labs will not wait on slow, random JavaScript when tokens per unit of compute fall by an order of magnitude or two. Parallel will, because it is completionist about those tokens. Search is still a real adjacency — nine out of ten LLM-for-work use cases, in his worldview, want the web — so labs can build or buy. On the day of the taping, Parallel announced a Google Cloud partnership: when you ground Gemini or other GCP models with web search, the options are Google Search or Parallel Search.

Human and agent interfaces to search are not the same problem. Humans rely on keyword search. Huang asks whether that is because people can only hold a few words in memory. Agrawal’s correction is blunter: we can hold a lot of words, we are just lazy. We want to type two incomplete words with a typo and hope Google guesses, or lean on autocomplete so we never type the third word at all. Models have fewer typos, better-specified and often longer queries, and leave the engine less to infer. It is a different class of retrieval problem.

Reed’s complaint is the pre-AI slop that occupies the top of Google for “best product for X”: affiliate pages a practiced human can smell from a mile away. How does an agent not get fooled? Agrawal starts with why those pages exist. Ask for a public company’s latest headline revenue and everyone in the room knows the authoritative number sits in an SEC filing, perhaps on page 73 of a PDF. If Google ranked that filing first, the PDF might take 3.5 seconds to load, and you would still have to grep your way to page 77. Humans will not do that work.

So the web grew a class of pages that load fast, put the number above the fold, are 99.99 percent right, and pull twenty facts out of a 300-page earnings report. You can call that slop, or you can call it SEO that successfully catered to a lazy human. Agents do not have to click around and fumble a PDF. Parallel takes an excerpt from the most authoritative place on the web and drops it into the model’s context window. The trade-off that created the middleman page is a leftover of browsers and protocols that never consistently pointed at the exact paragraph.

Swap ordinary search for Parallel, he says, and an agent typically uses under half the tokens, comes back more accurate, and finishes faster. Saved context is harder problems, or the same work cheaper; he treats that as compute allocation across the model, agent, and search layers. Huang asks where quality tops out. A billion-to-billion matching problem is a forever problem, Agrawal says, and the real question is when extra optimization stops being worth the squeeze — a day he is betting they never reach, because anytime you can matchmake 20 percent better than somebody else, that is an edge.

Agrawal used to build ads systems, and he says he intellectually loves them. Ads are why Google Search and Twitter could be free, and why so much web content is free. The mechanism is differential pricing: Google loses money on most queries and makes it up on some, with extraordinary margins, and Twitter had the same shape. Huang, who actually likes ads and shopping, jokes that she is subsidizing everyone else’s free information. “You are subsidizing all of us. Thank you,” he says. The core assumption is limited human attention converting into outcomes. If humans do not show up and agents do, that machine stops.

Publishers already see the crack. They will pay for SEO so more humans arrive, then block the agent that is acting on a human’s behalf, because that visit cannot be monetized. Subscriptions have the same hole: a thousand human visitors converting twenty subscribers is a known funnel; nobody yet has statistics on whether nameless agents subscribe or just take the content. For high-quality owners, the one available deal before Parallel was a head-of-market fixed-fee contract with a lab — some mix of training data, liability protection, and inference-time access. Most of the web cannot get that deal. Even the head, Agrawal argues, is on a broken model.

Inference might grow 7x this year and another 7x the next, on the order of 50x, while the contract does not. Nobody who signed a two-year deal believes their share will not shrink materially at renewal. He has been on both sides: he sold Twitter data to OpenAI. The properties of a fix, learned from ads, are efficient differential pricing on two axes at once — unique, scarce content costs more, and high-value work done with the same content costs more — done at scale, not as bespoke deals. A banker reading a page should pay more than Agrawal’s retired dad. The math that formalizes “how much incremental value did this source add?” is Shapley values.

Huang, who studied game theory, lights up. Shapley values ask how to split a pie that is bigger than the sum of parts so everyone still wants to collaborate; SHAP feature importances in machine learning use the same idea. A full calculation to award a publisher one dollar might itself cost several dollars, so Parallel estimates: drop a source, watch quality fall, test whether a cent of extra compute claws it back, and train models on those runs. They have already announced partnerships on that logic, and the original incorporation was literally Shapley Inc. — “a terrible name for a B2B product” they hid behind NewCo badges for six to eight months. Allocate two to ten percent of knowledge-work LLM inference to web data, he says, and the pool already beats every open-web data business outside walled gardens like Facebook and LinkedIn; with agent traffic compounding an order of magnitude a year, he puts meaningful checks 12 to 24 months out.

They landed on the name Parallel because they were doing a lot in parallel, and because they started visualizing a parallel web whose properties differ from the human one. Publish a page and you now have two audiences: what it should look like to a person, and how an agent should be able to read it. Reed is confident that earnings transcripts have already flipped — more people consume them through agents than by listening to the audio or reading the file. Agrawal says the same of Parallel’s API docs: customers are building with AI, so the primary reader is an agent, and that is how they test the docs. Cloudflare has reported AI page reads roughly even with human traffic, but he cautions that this counts crawlers and is not the same as search queries overtaking humans. That crossing has not happened yet.

He sketches three levels on the journey. Level one is where most products still live: a model given the same search tool humans already know, because the models were trained to use those tools. Level two is showing up in a subset of customers — multi-agent systems that wake subagents, familiar inside coding harnesses, and more exotic in long-running “AI scientist” setups that throw large amounts of compute and data at a hard problem. Level three is specific to the web: it goes from pull to push. Today an agent or subagent is told to go find this for me right now. In a couple of years, the pattern becomes call me if this happens.

The line he repeats with his team is a scheduling rule. If something can be done today, is worth doing, and GPUs are still available, they just do it; tomorrow is not for leftover chores. Tomorrow’s work is a reaction to a change: another agent finishing, a shift in satellite imagery, a scrap of customer commentary, a human with a new insight. One of those feeds is everything that changed on the web. Parallel’s role, if the vision holds, is to keep allocating compute across the whole web on behalf of every customer, continuously. His own background agents are bounded only by value versus spend. He admits he is probably overpaying a bit right now. It is not enough to care yet.