Parag Agrawal
Related people
Episodes
Notes
Sourcery · Ex-Twitter CEO Says We’re Underestimating AI Agents by 10,000x
-
How Should AI Agents Pay Content Owners?
Parag Agrawal says an AI agent should pay each content owner every time it reads that owner's content, and Parallel computes the amount with a model built for that. The payment is there so an owner has a business reason to allow the read instead of blocking the agent.
He describes it as his version of AdSense: easy, automatic, and differentially priced, so higher quality content is worth more to the owner. Parallel has started partnerships with content owners. His bet is that paying on each read is what keeps the web open.
An owner who cannot see how an agent visit becomes a business will shut the agent out. The per-read payment is the incentive he wants in place of that block.
-
Why Does Parag Agrawal Call Pressure Made Up?
Parag Agrawal calls pressure a made-up thing and does not feel it. During the Twitter acquisition he tracked his sleep score and was sleeping fine. If you know what you must do, you do it, and you do not have to feel pressured. The world can come at you. Knowing what you are capable of, with a slightly long-term view, makes everything a week or a month. You still have to perform every day, which is completely different from pressure.
The feeling he wants is insecurity. He has always told people he only wants jobs he would not hire himself for, so someone throws him into a challenge that on paper he is not supposed to be able to accomplish. Feeling insecure keeps him slightly humble and a li
-
Why Does Parallel Keep a Machine Version of Its Website?
Parag Agrawal says Parallel still keeps a machine version beside the human version because the company is building a parallel web for agents. As far as he knows, Parallel was the first website to use the metaphor. Three years ago, while the company was in stealth and the public site said almost nothing, a machine version was already live. The site still shows the human version, and a toggle switches it into the machine version.
The world is coming around to the metaphor, but the presentation layer almost does not matter. Parallel can transform the web even when a publisher does not make a site agent friendly. The product reads the human version and helps it become agent-friendly. Parallel d
-
Why Won't Models Alone Stop AI Injection Attacks?
Parag Agrawal says securing agents takes defense at every layer, and models alone will not solve it.
The web is a dangerous place. Anyone can post anything malicious anywhere, and a lot of that content will flow into an agent's context, so injection attacks are possible. Model builders have made very real strides in the last two years on not being very vulnerable. The whole world will rely on defense in many layers, and almost every layer has to obsess about the problem. The web layer has to focus on solving it before the problem reaches the models. Models then have to focus on it. App builders who build with models and with the web layer have to focus on it at their own layer. Dealing with
-
Why Should AI Agents Declare Themselves?
Parag Agrawal says good agents should declare themselves so other people can recognize them and treat them differently from bots that might be adversarial.
Most bad activity never shows up on the surface. An account that looks like a bot is harmless, because anyone can see it. The real problem starts in the margins, when people and the technical teams building the systems can no longer tell a human from a bot or an agent. Agent is the better word, because bot carries a more negative meaning, and agents are perhaps the good bots, mostly. Protocols are already being built to declare identity. The core challenge is whether an agent should inherit a person's full access or only part of it. Gett
-
How Do AI Agents Change Cybersecurity?
Parag Agrawal thinks the cybersecurity risks are real and that agents help both defenders and attackers. People who rush ahead to adopt agents for defending themselves will be fine, because attackers like to find the weakest link and there will always be a weakest link. The security risk for something highly sensitive, like the model weights of a frontier model, is a whole different conversation.
By and large, his intuition is that coding agents will be used aggressively to secure, but along the way a cyber event of consequence is going to happen every few weeks. The events will drive action and urgency in getting ahead on defense, and a bunch of influence is going to be spent on both attac
-
Why Won't Web Ads Last With AI Agents?
Parag Agrawal says web ads and per seat subscriptions are not durable once agents browse for people, because attention can no longer be monetized.
The older internet businesses made money from humans wandering the web. A publisher could put content out, watch users arrive, and convert some fraction of them into a monthly subscription, the way a newspaper does. Other businesses got paid when human visitors saw ads. Still others used affiliate offers or drove a transaction. All of the older models relied on humans random walking the web, opening pages, and taking the next action after the next action. Only some of the actions could be charged for. When agents follow the same path, the buying
-
How Much of the Web Do AI Agents Use?
Parag Agrawal says agents are using the web at 10 to 20 percent of humans. If his thousand X number is right, at least 5,000x or 10,000x growth is left from here. He treats the remaining growth as another way of saying we are very early with agents.
A lot of agent traffic is also people crawling for pre-training. When a website owner looks at automated activity versus human activity, automated activity is ripping. The 10 to 20 percent level is his best estimation of where we are right now. In the Bay Area and Silicon Valley, people are living in a little bit of a bubble. Most people's agents are not doing the number of searches his agents are doing, but they will.
-
Is a Thousand Times More Agent Web Use Too Low?
Parag Agrawal says one thousand times more web use by agents than by humans is too low. Three years ago he told people that agents would use the web one thousand times more than humans ever have, back when agents did not work and most people did not believe him.
An average person does on the order of ten or fewer Google searches a day. The agents he tracks run several thousand searches a day. He had treated one thousand times as the wild number for the pitch, and the deck had a section asking if Parallel was too early. Parallel was not too early. Across the last three years he was wrong on the other side.
He believed the idea because he had obsessed over agents searching the web for a long