Grit · 27 Jul 2026 · From the week of 27 July
Ex-Twitter CEO on Why AI Needs a New Internet | Parag Agrawal
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Kleiner Perkins partners Joubin Mirzadegan and Mamoon Hamid interview Parag Agarwal, former Twitter CEO and now founder and CEO of Parallel, which builds web infrastructure for AI agents. Agarwal covers his personal operating system: three priorities a day, deliberately dropped balls, raw Friday demos and a self-built AI agent. He also covers how Parallel manages token spend, why it is hiring engineers as fast as it can despite AI productivity gains, and its engineering principle that 'the answer to every problem is a model.' On go-to-market, he describes realising Parallel closed nearly every deal it entered but entered too few, which led Kleiner's sales operating partner Liam to join. His central thesis is that agents will use the web 1,000x or more than humans ever have. That will require new infrastructure and new business models that pay for open, high-quality content, and the biggest opportunity is entirely new work rather than cheaper old work.
For founders
- Agarwal argues that rising AI productivity is a reason to hire more engineers, not fewer, when the opportunity space is large: Parallel doubled from about 15 to 30 engineers in six to seven months and is still hiring as fast as it can.
- Agarwal's engineering default is 'the answer to every problem is a model': collect data, define a good output, train a model and put it in an improvement loop instead of writing code, which he says frees him from deciding which work is worth doing.
- Agarwal sets three proactive priorities each day or week, accepts that other balls will drop, and relies on important items resurfacing and on the team learning to catch what he drops.
- Agarwal recruited outside his own network by asking strong people for the best three people they had ever worked with, and deliberately kept more than half the team new; half of the six people who joined him on Parallel's founding team were people he had not previously known.
- On token spend, Agarwal frames the choice as which way you would rather be wrong, and chooses to over-spend rather than under-use models, while warning against measuring people by how many tokens they consume.
For revenue leaders
- A near-100% close rate can be the warning sign: a board member's observation that Parallel 'just not enough deals' revealed a pipeline problem, not a product problem.
- An outside sales operator who has seen many companies can calibrate how unusual your product pull is; Parallel's insiders had no reference points until Liam pointed it out.
- Parallel reportedly sold tens of millions of dollars of revenue with about four salespeople before building out GTM, and is now prioritising GTM, marketing and developer relations hires.
- Joubin Mirzadegan treats the solutions-engineering bottleneck of building custom demo environments as a target for automation with Claude skills, framing it as the difference between hiring ten more people and hiring three.
- Agarwal deliberately sells to AI-native products his team personally loves (Rogo, Harvey, Profound, Granola, Clay), because demanding, high-taste customers push the product forward and more follows from winning them.
What was said 33, most useful first
Agarwal is hiring engineers as fast as possible despite AI productivity gains, because hiring speed, building capacity and customer adoption multiply into outlier growth. Listen
He says it seems counterintuitive, but if there are many valuable things to build and customers need them, more engineers are better until you saturate the opportunity space. He attributes today's extreme growth rates in some companies to multiplying three factors: how fast and well you can hire, how quickly you can channel that capacity into valuable things, and how fast customers adopt them. He has never once concluded he needs fewer people because everyone is more productive.
“I have never once sat down and been like, oh, Everyone's going to be so much more productive. So maybe I need fewer people.”
Listen to the episode Hiring & team building Link to this Report a problem
Parallel's core engineering principle is 'the answer to every problem is a model': collect data, define a good output, train a model and put it in an improvement loop rather than writing code. Listen
Agarwal repeats this internally. When tempted to write code for a problem, the team instead collects data, figures out what good output looks like, picks a model architecture, trains it and puts it in a loop to keep improving. He says the mindset removes the need for him to curate which work is worth doing and lets him hire versatile problem solvers. The goal is to push more of the system into learned models rather than code.
“You're tempted to write a bunch of code for it? No. Instead, collect some data. Figure out what a good output would look like. Figure out what model architecture is right. Train the model and put it in a loop so that you can keep improving the model.”
Systems once built as heuristics stacked on models and made end-to-end over years can now be built end-to-end from day one. Listen
At Twitter, and in self-driving at Waymo by his account, systems began as cobbled-together models with heuristics on top and became end-to-end trained as teams learned the problem and collected data. He believes companies no longer need to go through that journey slowly. At Parallel they are trying to jump straight to 'the fourth step', which he says puts them on a path of rapid self-improvement.
“you can shoot more end to end from the very beginning at scale, and I think doing so puts you on this path of rapid self-improvement.”
A board member's observation that Parallel closed almost every deal it entered showed that its problem was too few deals, not winning them. Listen
At a board meeting early in the year, Kleiner's Mamoon Hamid asked how often Parallel closed when it showed up; the answer was 'we tend to always close.' His reply, that they were 'just not in enough deals', hurt, and Agarwal called him the same day asking for help. Parallel had been recruiting a head of sales for some time, and Hamid got Kleiner's sales operating partner Liam to help.
“how often do you win? How often do you close? And we said we tend to always close and once like so you just not enough deals.”
Listen to the episode Sales process & deals Link to this Report a problem
Agarwal built a founding team beyond his network by asking strong people for the best three people they had ever worked with, and kept more than half the founding team new. Listen
Parallel started in stealth, so there was no inbound interest. For almost a year before hiring, he met many people and asked each a standard question about the best three people they had worked with, then met some of those. The founding team was him plus six others, and he deliberately made half of them people he had not known before. He says much of Parallel's later hiring has come through this multi-hop network channel.
“what are the best three people you've ever worked with? And I would go meet some of them over time.”
Listen to the episode Hiring & team building Link to this Report a problem
Agarwal predicts that agents will use the web far more than 1,000x as much as humans ever have, which will require new infrastructure and new business models. Listen
About two and a half years ago he wrote down that agents would use the web 1,000x more than humans, and he now thinks that undershoots. From his experience building systems at that scale, he says no existing system is architected for it. He also argues that the web's three-decade bargain, in which content is published openly in exchange for distribution and monetisation, is breaking: the web is already starting to close up. Incentives therefore need to reward high-quality, unique content for it to stay open.
“agents and AI's will just use the web a lot more than humans ever have. And I wrote down a number which is, like 1000 x”
Listen to the episode Strategy & market Link to this Report a problem
Agarwal predicts the next wave of AI value comes from entirely new work rather than doing existing work faster or cheaper. Listen
In what he calls a semi-hypothetical example, a PE firm that once used gut feel to narrow buyout targets to four or five could use Parallel and models to run near-exhaustive 'simulations' across criteria before choosing. Because capital and people limit how many deals the firm can do, picking better creates enough value to justify heavy token spend. He calls this the next year of growth: 'work that was not happening yesterday.'
“It's going back to what does it take to make it 1000 x, right? It is work that was not happening yesterday.”
One of the Kleiner partners is tracking whether token spend per worker, which they estimated at about a tenth of a salary today, rises to half or more of a salary. Listen
The investor, whom the transcript does not identify, says they ask most of their portfolio CEOs about token spend to understand the curve of usage against falling costs. They floated that spend might be around one tenth of a person's salary now and asked whether it goes to 50% or one times a salary. They said they don't know where it lands.
“if it's right now, maybe like one tenth of a person's salary, does it go to 50% of someone's salary to one times a person's salary? And I don't know where it sort of lands”
Listen to the episode Metrics & finance Link to this Report a problem
Agarwal picks three proactive priorities each day or week and deliberately accepts that other balls will drop. Listen
Parag Agarwal, CEO of Parallel, says he is 'extraordinarily comfortable dropping balls'. Each morning he writes down the three things that must get done that day or week, shaped by what he has been reading and thinking about, and makes sure those happen. Reactive work fills the remaining time, but it is not allowed to swamp the proactive list. He bets that the team evolves to catch the balls he drops, and that anything important enough will bounce back to him.
“every week, every day I have like a few things that I care about or that matter. And I'm just going to do those things and there'll be a bunch of other balls that are less important that I will drop.”
Listen to the episode Leadership & culture Link to this Report a problem
Agarwal separates reading email from responding to email so that important items can rise into his top three without him triaging every message. Listen
He reads all his email in one work cycle and responds in another. He does not look at his inbox when choosing the day's three priorities: if something was important enough as an upside opportunity or downside protection, it should stick in his head. He says he is not trying to be 'some machine trying to prioritize every decision because that's exhausting too.'
“I separate out reading and responding to emails, the two different work cycles for me.”
Listen to the episode Leadership & culture Link to this Report a problem
Parallel's Friday-evening demos are for raw, unfinished work, so the CEO sees progress his priority list would otherwise hide. Listen
According to the host, Parallel holds demos on Friday afternoons from 4:30 to 6. Agarwal says they are explicitly not for packaged features shipping on Monday; they are for 'I was doing this this week, it's not done, here's why I'm excited.' Without them, this work would never bubble into his top three and he would only learn of it once it was packaged. Over the weekend the demos trigger excitement or paranoia about missed opportunities, which he has to boil down into something actionable by Monday.
“Friday evening demos are for. I was doing this this week. It's not done. It's raw. Here's why I'm excited about it. And you should all know about it.”
Listen to the episode Leadership & culture Link to this Report a problem
Agarwal built a personal AI agent with access to his Slack, Granola, email and the open web; it infers his priorities and pushes back on the balls he drops. Listen
He built it partly as an experiment to see how much a machine working on his behalf spends on Parallel's web access versus internal data. It knows his written top three priorities, acts proactively and is 'paranoid on my behalf,' which makes him more deliberate about which balls he drops. The host said Agarwal had spent $378 on tokens by 1 p.m. one day. Agarwal calls the original version 'outrageously inefficient'; after optimising it he expects to spend about $100 to a couple hundred dollars a day.
“it makes me feel more secure that I am not dropping the important balls because it is also paranoid on my behalf. It is inferring my priorities.”
He would probably let any Parallel employee who can drive value spend on a personal agent, but he has not done the math and does not yet know how to productise it. Listen
Asked whether everyone at Parallel could spend, say, $500 a day on such an agent, he said 'Probably, yeah. I haven't thought it through.' He cautions that nothing is useful out of the box: you have to mold yourself around it and mold it around you, which takes real effort. For anyone who does that work and drives value, he said $50 a day is fine.
“I haven't thought it through. I haven't done the full math, but yeah, if once useful. Totally.”
Much token maxing is not valuable and that people should not be measured by their token consumption. Listen
He says a lot of token maxing, including some of his own over recent months, is not productive. The cost he worries about is less the tokens than the time and energy spent, plus 'AI psychosis'. Some people are very good at using heavy token use to accelerate themselves and many are not, so token volume is the wrong metric.
“I think it's really important to not measure by token maxing.”
Agarwal's framework for AI spend under uncertainty is to decide which way you would rather be wrong; he chooses over-spending. Listen
Because no one knows the right level of token spend at any given time, he asks whether he would rather be wrong by being too conservative or by spending too much time and tokens with models. His personal choice is to err toward token maxing. He says you will never be exactly right, but knowing which way you prefer to be wrong tells you what to do.
“if you know which way you'd rather be wrong, that tells you what you should do.”
At Roadrunner, token spend is consolidated and tracked by model but not capped, because limiting the best engineers' agent experiments feels like a mistake. Listen
Joubin Mirzadegan says his company Roadrunner ran 'let it rip' for about six months, then moved all spend onto one credit card to see where it went, and then tracked which models it went to. They track spend but enforce nothing. The one rule of thumb is that an engineer spending their salary on tokens 'better be pretty good.' He would rather pay more now so his best engineers can stay on the bleeding edge, for example chaining five agents together.
“the short answer is we track it, but we don't enforce anything today.”
The real business value of AI agents is people expanding their span of ownership, not doing the same work faster. Listen
His examples: backend engineers doing more frontend work, frontend engineers changing APIs, and people using agents to run security reviews or push optimisations they would otherwise never have done. He sees true value in people taking more end-to-end ownership and pushing their own learning curves.
“it's actually people expanding their span of influence and ownership on the product. That's where I think there is true business value being generated in my mind.”
Mirzadegan pushed his solutions team to automate custom demo-environment setup with Claude skills instead of adding headcount. Listen
Prospects of his company Roadrunner want demos built on their own messy SKUs, and setting up each environment took his head of solutions architecture a couple of days of 'hand-to-hand combat.' He asked what it would take to cut that to a couple of hours, and suggested embedding the needed engineering skills into Claude to try to automate it, while acknowledging it might fail. His framing: with 15 customers at once the manual process breaks, so the choice is hiring ten more people or automating and hiring three.
“either we go hire another ten people or we figure out how to automate this process and we can hire three.”
Listen to the episode Sales process & deals Link to this Report a problem
Wherever a task can be measured cleanly it can be put in an automated model loop, so engineering work becomes increasingly meta. Listen
At Twitter, many engineers improved recommendation systems by intuition, running A/B tests, finding features and building data pipelines. Agarwal says much of that can now run in automated loops wherever there is a clean way to measure quality, and big parts of Parallel's system are such loops. Engineers' work shifts to finding feedback loops they trust and making them measurable, so they need to look under the hood less.
“wherever there is a nice, clean way of measuring things, right, you can automate a lot of those things”
Agarwal's lesson from Twitter is to rebuild systems at each order of magnitude of scale, because building for four years out means shipping nothing. Listen
At Twitter he says they had to rebuild entire systems roughly every year because the systems weren't built for the next order of magnitude. He has carried the same lessons to Parallel, which he expects to become meaningfully larger in scale than Twitter: design pragmatically for the current scale and know when to rebuild.
“if you build for like four years out, you don't ship anything. And so you have to constantly keep evolving for scale.”
An outside sales operator gave Parallel the calibration to see how unusual its market pull was. Listen
Agarwal says the team had not internalised how strongly the market was pulling its product, because insiders had no reference points. Liam, arriving as an outsider, looked at things objectively from week one and told them how unusual their numbers were. Agarwal says this gave the team more confidence and direction.
“it was him telling us about how unusual it is, what we were seeing on the inside, because we had no calibration points. He did.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Mirzadegan relays Liam's assessment that Parallel had sold tens of millions of dollars of revenue with about four people, 'in spite of themselves.' Listen
Liam, Kleiner Perkins' sales operating partner, had spent a lot of time inside Parallel and talked to all its reps before telling Mirzadegan he was at least 50% interested in joining. Mirzadegan says salespeople have a good nose for whether a product will sell, and Liam later joined Parallel. Mirzadegan tells his own team that the measure of doing a good job is a founder trying to hire you.
“they've sold like tens of millions of dollars of revenue with like four people and like it's in spite of themselves.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Agarwal built a months-long relationship with Graham, previously head of sales at Windsurf, before hiring him at the point Parallel could fully use him. Listen
Mamoon Hamid introduced them early as a GTM sounding board, before Parallel had any salespeople. They spoke for several months, and Graham believed in the company enough to want to invest. Agarwal hired him once Parallel reached the point of scaling the GTM motion. He says these small things compound into deep partnerships, and that you are building a company of 'believers and missionaries' in every function.
“ultimately you're building a company of believers and missionaries, and no matter what you roll, what function you're in.”
Listen to the episode Hiring & team building Link to this Report a problem
Parallel's hiring priorities have moved from engineering-only to GTM, marketing and developer relations as it builds the company around the product. Listen
Agarwal says Parallel would hire across every function but proactively focuses on a few. Engineering is always one; GTM, and newly marketing and developer relations, are the top additions. He describes the phases as building the technology, launching the product last year, and now building the company around it.
“Always engineering. Now GTM and the new ones are marketing. Developer relations are the top new ones on my list where we just need to do more.”
Listen to the episode Hiring & team building Link to this Report a problem
AI products create an 'all-knowing' expectation, which makes web access mandatory: without it, models feel unusable for about 75% of tasks. Listen
Early ChatGPT had no web access and still impressed people. Agarwal says that for 75% of what people now do with these products, a model without web access feels unusable, because the interface implies the model is smart and all-knowing. He gives coding agents as another example: users once pasted documentation links in by hand. He positions Parallel as infrastructure that sits next to every model at inference time.
“these models present interfaces and a mental model, which is that they're both smart and all knowing Once you build that mental model into a product, you must use the web”
Content that was not worth crawling for human search, such as deep-web portals without URLs, becomes valuable when the goal is to fill an AI's context window. Listen
Traditional search indexes optimised for 'ten blue links' skip content you can only reach by navigating or searching a portal, because there is nowhere to land a person. Parallel builds crawlers for that content because its job is to bring content into an AI's context window. Harvey uses Parallel's search so its output is grounded in authoritative public documents, and needs both completeness on hard-to-crawl data and strong ranking. Agarwal says Harvey's demands push Parallel's capabilities.
“historically people, when it's hard to crawl two it's less valuable to crawl because you can't actually land a person on the content. In a world of eyes. Our job is to bring the content into the context window of an AI.”
Listen to the episode Strategy & market Link to this Report a problem
Parallel deliberately pursues AI-native products its team personally loves as customers, because demanding, high-taste builders push the product forward. Listen
Customers cited on the show include Rogo, Harvey, Profound, Granola and Clay. Agarwal says the team believes it has good taste and that products built by high-taste people are extremely demanding about what they use. His core belief is that if those products love building Parallel into their own, 'there's a lot more that follows from that.'
“if we can be a product that they love putting into their products to power them. There's a lot more that follows from that.”
Listen to the episode Positioning & marketing Link to this Report a problem
Parallel doesn't need to expand its market, because any model used for work must be given the web; the question is only through which product. Listen
He describes two routes. Enterprises can buy Parallel's API directly and connect it to internal data through MCPs to build their own workflow automation. Or they can buy from AI-native vendors that embed Parallel. Over the next year he expects a class of top AI-native agent products and a class of large enterprises to incorporate Parallel 'if we do our jobs right.'
“I don't think we have to expand the market like we are.”
Listen to the episode Strategy & market Link to this Report a problem
One of the Kleiner partners argued that companies growing below the new AI-era rates are good companies, not great ones. Listen
Responding to Agarwal's point about growth factors multiplying, the investor said companies with the tools, wide blue oceans and capital are growing at rates never seen before. By that standard, slower-growing companies are merely good. The transcript does not identify which partner said this.
“if you're not growing up those rates uh, you're a good company. Not a great company.”
Listen to the episode Strategy & market Link to this Report a problem
Parallel roughly doubled its engineering team from about 15 to 30 in six to seven months and has 40 to 50 people in total. Listen
Agarwal gave these figures for Parallel, a company building web infrastructure for AI agents that launched its product the prior year. He frames the stage as moving from building the technology and product to 'building the company around it.'
“In about 6 or 7 months, we've doubled.”
Listen to the episode Hiring & team building Link to this Report a problem
Agarwal considers every exceptional person he meets as a potential hire. Listen
He says it is very hard to believe someone is exceptional, know you will eventually need that kind of talent, and not picture recruiting them. He describes walking the world looking for great people who believe in the mission.
“every special person you end up talking to as a company builder, you think about hiring if they are the right person.”
Listen to the episode Hiring & team building Link to this Report a problem
Agarwal dates the take-off of Parallel's demand to about nine months earlier, when people started getting real value from models in work contexts. Listen
He says Parallel grows when new categories of work start deriving value from models. In his observation this began in some areas the previous summer, about nine months before, and has been exploding since. Frontier agent builders across categories are now incorporating Parallel's APIs.
“that really started happening in my observation last summer in some like nine months ago in some areas, and has been exploding since then.”
Listen to the episode Strategy & market Link to this Report a problem
When Agarwal asked to hire Kleiner's sales operating partner, Mamoon Hamid's response of 'if it's right for you, it's right for me, but don't do it' convinced Agarwal he had picked the right investor. Listen
Agarwal called asking a board member's firm for its own operating partner a hard conversation. Hamid first said Kleiner is all about founders, so if it was right for Agarwal and Parallel it was right for him. He then argued against it and negotiated compromises to try to keep Liam. Agarwal says he needed both responses: without the pushback he would have wondered why Hamid wasn't fighting for Liam. He left feeling validated both in trying to hire Liam and in choosing Hamid.
“He's like number one. We're all about founders. And so if it's right for you and parallel, it's right for me. Second, don't do it. And I kind of needed to hear both of those things.”
Listen to the episode Fundraising & investors Link to this Report a problem