Parag Agrawal Bets on Search Technology for AI Agents

Parag Agrawal bets that AI agents will use the web 1,000 times more than humans do — and that the infrastructure for that traffic does not exist yet.

Twelve years after joining Twitter as a junior engineer, and two years after steering the company through one of the messiest transitions in tech history, the former chief executive is staking his second act on that idea.

It is the thesis behind Parallel Web Systems, the company he founded and now runs as CEO, and the through-line of a long interview on the Knuckle Up with Nakul podcast published earlier this month.

Parallel builds search technology for autonomous AI agents rather than human eyeballs. It has grown roughly fourfold to about 120 people in under a year, Dealroom News reported. Backers include Kleiner Perkins and Sequoia — two of the firms that for decades decided which infrastructure bets mattered.

The timing matters. Agent traffic is becoming a measurable share of web requests, and the companies that own the current access layer — Google’s search index, OpenAI’s crawlers — have business models built around humans, not machines that read 10,000 pages a second. Agrawal’s argument: nobody building a frontier model has the incentive to make search work for everyone else’s agents. Whether that gap is real is the entire commercial case for Parallel.

From physics gold medalist to the ads team

The interview opened with what Agrawal calls a default position for his career.

Asked why he would start a company at 40, his answer was simple.

Video: Knuckle Up with Nakul

“40 is young.”

The resume supports the pattern. A gold medal at the International Physics Olympiad at 17. Top of his computer science class at IIT Bombay. A Stanford PhD. Yet he insists he was never an ambitious child — happy-go-lucky, drawn to maths and physics only because he and an exceptional peer group found them fun, with no role models or idols.

Ex-Twitter CEO on Why AI Needs a New Internet — Kleiner Perkins

Academia lost him to its signal loop: peer review was slow, and the feedback low-signal. So he joined Twitter as a junior engineer on the ads team. From there the numbers get extreme. The team grew from around 10 engineers to hundreds.

Infrastructure scaled from four computers to 4,000 in roughly 18 months. Revenue went from almost nothing to a couple of billion dollars in two to three years. He learned distributed systems by building the ad stack that paid for all of Twitter’s other engineering.

Hiring without precedent

His rise from junior engineer to distinguished engineer to CTO to CEO within a decade, he says, was not talent hoarding but the opposite: surrounding himself with more capable people and working problems from the outside in. When recruiting stalled, he got on a plane — flying to India to hire IIT undergraduates with no precedent, printing offer letters the day of marathon six-hour interview sessions.

Then came the job itself. Once in charge, he stopped reshaping himself to fit the company and adapted the company to fit him — his personality, vision and mental model dictated how the business was organized and hired. A CEO with a five-year view has to transform the business and own themselves.

Running Twitter in dual reality

The most operationally interesting stretch is his account of running Twitter during the contested takeover. Two contradictory truths at once: staying committed to closing the deal while operating the company as if it would never close. He credits lawyers with influencing almost every decision in that period — and strong leadership for keeping the product shipping throughout. Satellites kept flying; code kept merging.

That is a rarer sentence than it sounds in 2026. A company that kept a normal release cadence through a hostile ownership fight is a genuine case study in organizational stability.

Search for machines, not people

Agrawal built agents as a hobby after leaving Twitter, “when they didn’t really work.” The hobby became an obsession. The obsession became Parallel.

The reasoning is arithmetical, not mystical. Let agent adoption continue its trajectory and agent traffic will not merely rival human traffic in server logs — it will exceed it by three orders of magnitude. That web needs new retrieval built not for clicks but for machine consumption. Parallel is starting with search built specifically for AI agents, and says it plans to compensate content owners so the web stays open.

Why not Google, why not OpenAI

Agrawal makes the same call as most players: neither Google nor the frontier labs will build search for everyone’s agents. Google’s revenue model is optimized for human attention and advertising. The frontier labs need web access to power their own models but have no product-level incentive to route other people’s agents through it.

An independent player — one whose entire revenue depends on being a neutral access layer for third-party agents, with content owners compensated on the side — is, in his view, the only configuration that keeps the web open.

Two versions of the bet

In one, Parallel is a rational bet on an underserved infrastructure layer, backed by firms that have seen this movie before. In the other, agent traffic never reaches the 1,000x figure and the retrieval market consolidates inside the labs doing the agent work. The honest answer to who’s right is a table only 2027 can produce.

But the more interesting detail in the interview isn’t agentic. It’s “I only like jobs where I think that I am slightly unqualified for doing that.”

“40 is young,” Agrawal said.

Photo of author

Technology Desk

Technology Desk at Archyde.

New SARS-CoV-2 Variant PJ.2.1 Neutralized by XFG-Adapted Vaccine Boosters