AI Agents Could Drive 10,000x More Web Activity | Former Twitter CEO

Summary

• Agrawal sees 5,000–10,000x further growth in agent web use if his original 1,000x thesis holds. He estimates current agent usage at 10–20% of human usage, distinguishing it from crawler activity and headline bot traffic. His own agents already make several thousand searches daily: “1000x is too low.”
• Parallel’s growth benchmark is agent inference, with 10X as Agrawal’s current planning approximation. He expects the world to spend one to two GPUs on web search for every 10 GPUs spent on model inference. Capacity involves specific hardware, locations and timing; some GPUs must sit within five milliseconds of Parallel’s web index.
• Parallel sells search quality, model independence and lower inference costs. The introduction reports a $100 million Series B at a $2 billion valuation, while Agrawal says headcount grew from close to 30 to about 120 during the year. He claims higher-signal search lets models do “half the number of searches,” consume fewer tokens and finish faster.
• Consumer-agent retention depends on fulfilling an unusually simple product promise. Agrawal expects good retention because messaging an agent requires little onboarding, but acknowledges having no insider retention data. He rejects the suggestion that OpenClaw’s concept failed: setup, continuous maintenance and early model limitations constrained accessibility.
• Agrawal argues that agent browsing breaks attention-based web economics even when purchases survive. “You can’t monetize attention anymore,” so he calls advertising and per-seat subscriptions non-durable. Parallel is building payments to content owners for agent reads, with models determining compensation—a proposed “AdSense” for this web.
• Agent access requires commercial alignment as well as identity and security. Agrawal wants useful agents to identify themselves and inherit their owners’ permissions, yet says identification currently invites blocking. Security remains a separate, layered responsibility; he expects consequential cyber events “every few weeks” as attackers and defenders both spend inference.
• Agrawal’s next demand thesis is that “the web will go from pull to push.” Today’s always-on agents remain mostly reactive; he expects agents to monitor changes and act on users’ explicit and implicit needs. Across a trillion-page web with growing machine-generated noise, the constraint becomes allocating compute to actionable changes.

Deep dive

1. Agent web use remains early despite the traffic headlines

• Agrawal’s original pitch was that agents would use the web “1000x more than humans ever have.” Now, “1000x is too low”: an average person makes around 10 or fewer Google searches daily, while his tracked agents make several thousand. He puts himself roughly, but slightly below, 1,000x. Most people are not there; reaching that point might take a few years.
• O’Shea cites coding agents reaching 60% of traffic observed by Vercel. Agrawal cautions that automated traffic includes pretraining crawlers, and agents search more than humans to accomplish equivalent work. His best estimate puts agent web use at 10–20% of human usage. If his original 1,000x thesis is right, that implies another 5,000–10,000x from here.
• O’Shea’s pushback—will smarter agents eventually search less? Agrawal agrees, but rejects scaling old search infrastructure 1,000x. His best guess is that Google will do roughly 10 billion searches daily; multiplying that infrastructure a thousandfold is not something he thinks they can build. Parallel instead uses GPUs to underpin search indexes, historically CPU- and memory-bound, and distills ranking intelligence into smaller models.
• His conviction came from experimenting with BabyAGI and AutoGPT using models such as “GPT-3.5 or whatever.” They did not work, but the direction seemed predictable: better models running longer, with models becoming smaller and cheaper at constant capability. He did not predict coding agents would lead adoption or that the CLI would become their form factor.

2. Inference sets the growth hurdle and the capacity problem

• The introduction reports a $100 million Series B “last month” at a $2 billion valuation. Agrawal’s benchmark is inference: “for every 10 GPUs the world spends on model inference, we’ll spend one to two on web search.” His current growth approximation is 10X, higher than last year’s answer. He wants Parallel growing alongside or faster than agent inference, while questioning how many years 10X can persist.
• Parallel rents GPUs and has experienced multiple capacity scrambles over three months. Availability means the right GPU, quantity, location and timing; some must be within five milliseconds of its web index. For certain large clusters, he also asks whether the physical capacity can be committed for long enough. Agrawal calls every hyperscaler and neocloud. Availability gets established partly on iMessage, then finance teams handle paperwork. So far, GPU shortages have not forced Parallel to turn away a customer.
• Parallel’s BrowserBase integration, which shipped that day, powers search in BrowserBase. It illustrates the distinction between reading and acting: Parallel’s APIs search and access the web before agents use browsers to log in, buy or write. Agrawal’s “cloud for agents” combines inference, perhaps a sandbox, web search, browsers and API actions, personal or company-data connectors, plus observability, feedback loops and security.

3. Retention and enterprise adoption depend on delivering the promise

• Agrawal expects consumer agents such as Instinct and Muse to retain users well, while acknowledging no insider retention information. Messaging provides natural onboarding; the challenge becomes whether the product delivers. His test: “can it actually deliver on this very simple high level promise often enough to become a part of your life?”
• Asked whether OpenClaw hit a wall, he disagrees: “OpenClaw was right in all of the ways.” Most people cannot set it up and continuously adapt it as capabilities change. Early performance also depended heavily on a strong model such as Opus. Managed products can improve underneath users without requiring them to understand what changed.
• Early customers Harvey and Notion pushed quality; voice agents now push latency. Search needs roughly 150–200 milliseconds, versus roughly 600–700 milliseconds for human search. Expectations harden quickly: Agrawal bets disabling ChatGPT’s web search today would disappoint users on 90% of requests, even with a new model. “You will feel like a caveman.”
• Headcount grew from close to 30 at the year’s start to about 120, largely as Parallel built go-to-market capacity over the last six months. Agrawal gives three buying triggers: bundled search fails demanding work; customers need independence across proprietary and open-weight models; or scaled workloads require better economics. He says high-signal results enable “half the number of searches,” fewer tokens and faster completion.

4. Keeping the web open requires payments and inherited permissions

• Agrawal’s economic argument starts with humans “random walking the web.” Publishers monetize subscriptions, advertising, referrals or transactions along that journey. With agents, “the buying action still remains. Every other action breaks.” He calls advertising and per-seat subscriptions non-durable, making incentive alignment with content owners part of the infrastructure problem.
• Parallel has built a system that runs a model to compute what to pay each content owner whenever an agent reads its content, and is establishing content partnerships. Agrawal describes a version of “AdSense” with “easy, automatic differential pricing.” The bet is that compensation for quality content can keep the web open, replacing publishers’ incentive to block agents with a reason to welcome them.
• Asked whether bot detection still makes sense, he distinguishes crawlers, useful agents and malicious bots. He has “two contradictory feelings”: good agents should declare themselves, yet declarations can trigger blocking before incentives align. His Instinct and Muse cannot access many sites he can. He wants all agents, including those outside large labs, to inherit their owners’ permitted access and actions.

5. Security remains a continuing contest at every layer

• Agrawal calls cybersecurity risks real: agents help attackers and defenders. He thinks organizations adopting defensive agents early will be fine because attackers seek the weakest link; frontier-model weights present a separate, more sensitive problem. He expects aggressive use of coding agents for security, consequential cyber events “every few weeks,” and substantial inference spending on both attack and defense.
• His Twitter analogy is a “knife fight”: attacks can be contained, almost never eliminated, and the hardest activity is what defenders cannot distinguish. Web content can flow into agent contexts and enable injection attacks. Models have made real strides over the last two years, but “you can’t say, ‘Oh, models will solve it.’” Parallel, model builders and applications must each defend their layer; occasional failures will still escape.

6. Proactive agents change search and compute allocation

• Agrawal’s current obsession is that “the web will go from pull to push.” Instinct, Muse and Dot show glimmers, but remain more reactive than proactive despite being always on. He expects agents to understand explicit and implicit needs, set high-signal, high-precision, high-recall alerts across the web and determine when changes require action. Falling intelligence costs could make earlier alerting concepts work.
• The allocation problem is a trillion-page web, much of it junk, with many pages changing daily. “No matter how cheap you make compute, we don’t have enough compute to process everything.” Resources must prioritize actionable changes. Agent-generated publishing increases the rate of change and potentially the noise, making signal extraction more demanding.
• O’Shea distinguishes GEO for chat from AEO for agents; Agrawal initially treats them as equivalent, then accepts her distinction. Parallel has offered human and machine website views since its early stealth period. Still, he considers presentation secondary: “we can transform the web” even when publishers have not made their sites agent-friendly.
• Parallel designs documentation and onboarding around a developer accompanied by a coding agent. Customers can paste a prompt into their agent; the human authorizes access, adds a credit card and accepts terms. The agent can handle the rest. Agrawal sees designing the entire customer journey around agents as fundamental.

7. Internal expectations drive Agrawal’s operating philosophy

• Agrawal credits his wife, a physician and Andreessen Horowitz investor, with raising his ambition beyond what he thought possible. His childhood peers were children of nuclear scientists in India, surrounded by expectations of becoming the best. He connects that environment to hiring: “If you don’t hire outliers, you don’t get outliers. If you hire only outliers, you get only outliers.”
• Asked how he withstands pressure, he replies, “What pressure?” His sleep stayed fine through the Twitter acquisition; he distinguishes pressure from performing daily. “I only want jobs which I wouldn’t hire myself for”: insecurity encourages humility, a little paranoia, curiosity, effort and seeking help. Excepting his wife, he says his expectations have always exceeded others’—internal standards drive him more effectively than external ones.