A lender has extended credit to 10,000 small businesses. Every month, a human process works through that portfolio to judge whether its risk is rising or falling, leaning on information scattered around the web. Parag Agrawal, describing customers of his company Parallel Web Systems on Sequoia's Training Data podcast, said a developer handed that routine to an agent — software that plans steps and uses tools on its own. The review now runs weekly rather than monthly. The number of web searches behind it, on his estimate, multiplies by hundreds of thousands, perhaps a million.
Agrawal ran Twitter before its sale to Elon Musk and now builds search and indexing technology aimed at agents rather than people. He said Parallel started from a bet that agents would search and use the web a thousand times more than humans ever have, which would require both new technology and new business models alongside it.
The multiplier stack
The smallest multiplier sits inside ordinary use. A typical search agent, Agrawal said, will run somewhere between five and twenty searches even when it answers within a few seconds, "because why not." Write a prompt to ChatGPT on a middle setting and something in that range happens behind the reply; dial the effort up and it runs hundreds or thousands of searches. One human action, in other words, already turns into roughly an order of magnitude more searches than the person performed.
Parallel's early products chased bigger multipliers than that — cases like the credit portfolio, where a developer programs the work once and a process a person used to run monthly can run weekly, producing a dashboard and a collection of action items for someone to look at.
Agrawal's own desk shows the pattern in miniature. Asked whether he uses anything to prepare for meetings, he described an agent product plus custom agents that draw on Parallel's internal data together with web data through Parallel's own APIs. He wrote one prompt to build such an agent. It now does tens and hundreds of web searches for every meeting he has, and each new agent he builds for a new use case multiplies that again.
What stops the multiplication, in his account, is arithmetic rather than technology: background agents will be deployed to the extent that the extra compute produces incremental value. He does not claim to have reached that discipline himself. "I don't think it's rational right now," he said of his own spending — probably more than he should be spending, but not so much that he cares yet. Some rationalization, he expects, is coming for all of these agents.
Still early, by his own count
Asked whether the web now carries more agent queries than human ones, Agrawal said he does not think so yet. Three years ago, before ChatGPT, he would have guessed he ran 20 or 30 Google searches a day. Across all his agents today he reckons the figure is 100 to 1,000 times that, and possibly more than 1,000 times if he counts work at his company that is not assigned to any human. "We are the outliers rather than the norm," he said, adding that a step outside the industry's bubble finds people who have not heard the vocabulary at all.
He offered one outside number with his own caveat attached: he recalled seeing a figure he thought came from Cloudflare's monitoring of web traffic, showing AI traffic roughly level with human traffic in page reads. Page reads are not searches, he noted, and that count probably sweeps in crawlers and other things.
From pull to push
Agrawal sorts the journey into three levels. At the first, people build simple agents that treat the web as a tool, because a web search tool is the familiar thing to hand a model trained on the tools humans use. By and large, most work sits there today.
At the second level, a subset of customers builds more sophisticated multi-agent systems: sub-agents that wake each other up or orchestrate work in other ways. The most familiar version is sub-agents inside coding agent harnesses, but Agrawal pointed to others — an AI scientist, for instance, running very long jobs that throw large amounts of compute and data at a hard problem.
Both of those modes are requests: an agent tells a tool or a sub-agent to go find something now. The third level, which Agrawal places a couple of years out, turns that around. The web goes from pull to push. Instead of a point-in-time need, a customer leaves a standing instruction — as he put it, "call me if this happens so that my agent can do some work or a human can do some work."
His reasoning starts from a line he uses with his team. If something can be done today, is worth doing, and there are GPUs available, it gets done today; nobody postpones work for its own sake. So what is left for tomorrow is work triggered by change — another agent finishing some compute, something visible in satellite imagery, a piece of customer commentary, a human arriving at a new insight. One of those feeds, he said, will be everything that changed on the web, which would let Parallel allocate compute across the entire web continuously on behalf of all its customers.
Writing for a reader that is a program
That future reader is already shaping how things get published. Parallel's name, Agrawal said, came partly from picturing a parallel web built for AIs, where publishing means addressing two audiences: what a page should look like to a human, and how to make sure agents can read it too.
One of the hosts pushed the point further, saying he felt confident that more people now consume earnings transcripts through agents than listen to the audio or read the transcript itself, and that a public company chief executive today should make sure remarks will be transcribed and interpreted correctly by the agents, not only by the people theoretically listening.
Agrawal said the same holds for the documentation Parallel writes for its own APIs. Its customers are building AI solutions and using AI to do it, so in nearly every case it is their agents reading the docs and the code rather than humans fumbling around a docs page. "For us, the primary audience is an agent," he said, "and that's how we test our docs."
The conversation is available on the episode page.