Parag Agrawal. Parallel.

Nakul Mandan:

Google has the compute, the talent, the background, the index. Why isn't this Google's war to win?

Parag Agrawal:

I think you should invert the question. You had lawyers on almost every decision knowing that this was looking like a contested moment. Credit to the leadership team there. Twitter continued operating well. We shipped a lot of valuable things in that period. That I'm proud of. The deal isn't closed until it's closed. And so you are now operating in this weird dual reality of we are committed to closing this deal and we'll put in every effort in the court of law to make sure it closes. And simultaneously, we are operating like it won't. I could keep these two contradictory things in my mind. It is very hard to have thousands of people operate in this sort of weird dual reality. Me, I'm in this camp of like the token maxer camp, and then there is the responsible camp, which is like, oh, I need to build a reliable system in the presence of these token maxers.

Nakul Mandan:

Why be a founder at the age of 40?

Parag Agrawal:

40 is young. I only like jobs where I think that I am slightly unqualified for doing that.

Nakul Mandan:

What are you unqualified for now? Parag Agrawal has been exceptional in almost everything he's done in his career. As a student, he was an International Physics Olympiad gold medalist, ranked in the top 100 in the IIT entrance exam, graduated at the top of his CS class at IIT Bombay, and then a Stanford PhD. He joined Twitter as a software engineer, rose to become the company's first distinguished engineer, then CTO, and eventually CEO, all within a decade. And today he's the founder of Parallel, a multibillion-dollar company backed by Kleiner Perkins and Sequoia. In this conversation, we talk about what drove him to excel in every aspect of his career, his mental models, his operating principles, and the internet of agents he's building for. Knuckle up. Parag, welcome to the show.

Parag Agrawal:

Thanks for having me.

Nakul Mandan:

So, you know, before Twitter, before Parallel, You know, Physics Olympiad gold at 17 years of age, top of the CS class at IIT Bombay, Stanford PhD. So you really excelled in every aspect of your career. What were you as a kid like? Were you always very ambitious and competitive?

Parag Agrawal:

I don't think people would have described me as ambitious. People would perhaps early on during middle school, high school would have described me as happy-go-lucky. I think I was competitive towards high school, but I was competitive because I was with a group of friends who were all exceptional. And it's probably driven by curiosity and just wanting to be— because we found it fun. We found it fun to do math. We found it fun to do physics. And that's sort of what led me to—

Nakul Mandan:

it was your peer group.

Parag Agrawal:

And a peer group. I think I give a lot of— I attribute a lot to the people I was around.

Nakul Mandan:

Were you, were you looking up to something in those growing up years? Was, there Narayana Murthy or somebody like that, or in the US somebody that you looked up to?

Parag Agrawal:

No, there wasn't like a, I, I've— one, even back then and even till this date, like I am not the kind of person that like looks up at like role models or idols. Like that's just not my way of visualizing myself or the future. I think it was like much more day to day, which is like, oh, we want to go do this, or this problem sounds hard, or I'm with a bunch of friends where we're, talking to each other about problems we're solving. you end up in this— getting into IIT has this entrance exam that like is talked about, which requires real amounts of hard work. turns out, but if you're— if you are someone who actually enjoys that kind of problem solving and learning and is surrounded by many such people who also happen to be really good, it just— it's the cheat code to be able to do it with perhaps less pain than other people.

Nakul Mandan:

And then when you landed in America in 2005, you went to Stanford for your PhD straight away. And your plan was to be in academia, right?

Parag Agrawal:

Plan is probably an overstated word. I think the, journey I was on, like after Physics Olympiad, went to IIT, started doing computer science, hadn't really done computers before then, right? Like, so I went from math and physics to computers mostly because the smartest kids went to do computer science at the best institutions. So I was lucky enough to get in, got to be there. As a computer scientist, I was like more of a math person. So I did a fair amount of research during undergrad. So I really enjoyed solving hard research problems. and I came to Stanford in part because one, I wanted to just— like, it seemed like PhD was the path for me at the time to me, because like, I, my world wasn't broad enough to fully contemplate every other path. But I also knew that I didn't want to do the same kind of research I had done during undergrad. So I chose Stanford because it would be a fresh, clean start relative to other institutions I could have gone to where I would continue the same research which I was already doing. At Stanford, it's like, okay, we don't care who you are, show up and figure— you have the first year to figure out what research you're going to do. And I change fields. My journey is to get more and more applied every step of the way, perhaps a little bit impatient every step of the way. Like during undergrad, I was fine doing like just writing proofs, right? I wanted to start building things, and that started for me during my PhD.

Nakul Mandan:

And then what made the switch to Twitter happen? Because that feels like a departure from solving math problems to becoming a junior engineer at Twitter.

Parag Agrawal:

Yeah. So I think early on in my PhD, I realized that I didn't want to be an academic, but I still enjoyed research. So then my next— mental model became to do research, but within industry. At the time, there were all of these labs, like Microsoft Research, Google Research, Yahoo Research, AT&T Bell Labs, like on its, final legs. And so I thought like, okay, maybe you can have it all, right? Maybe you can keep doing research, but also have an impact and do real things. and so through the PhD, through a bulk of the PhD, I thought that is where I wanted to be. but towards the end, I got even more disillusioned with the sort of straddling the boundary of academia where I think, like, I got disillusioned by the peer feedback loop of other academics judge your work. Your output is a published paper of sorts. Usually it is judged by others who are also not entirely living in the real world, right? And so it felt like a slow feedback loop. It felt like a low signal feedback loop. And I was getting impatient to like do things which were— which felt more real and tangible. And so after doing like a lot of summer jobs at every industrial research lab, I gradually came to the realization like you should just like go build stuff. And then, so it was a pretty last minute thing. I think even like 6 months before ending up at Twitter, I would've said that I'm more likely to do research industry than a software engineering job. And I ended up at Twitter mostly because I was obsessed with the product, right? Like it felt early, but I was like, I was on it.

Nakul Mandan:

You were, you, a tweeter also, like, was it, were you a consumer or you were actively tweeting?

Parag Agrawal:

I was. Very active. I was like, it was, I, I think like working at Twitter made me tweet less over time, but yeah, I, I was a producer-consumer heavy duty prior to starting to work there.

Nakul Mandan:

And that was a phenomenal run, right? Like you, went from a junior engineer to their first distinguished engineer to the CTO to CEO. Or, if I'm correct, and Twitter also went through a few eras at that time. It started when you joined, it was still, it was a hit company, but still small. Then it became the public town square of the world. How were those early days at Twitter and how did it evolve in the various CEO journeys before you ended up taking over?

Parag Agrawal:

Yeah, so I started as like a junior software engineer. I went there. To work for sort of the new ads team. So Twitter was already at the time a product that was starting to work. The biggest problem on the product was bringing in— making it reliable, bringing in enough machines so that it would scale and you wouldn't see the, the, fail whale. But on the other side, as the product was scaling, they needed to ultimately be a monetization story. And people sort of started bringing in like a small team to go build that. So, and that team essentially worked a little bit separate from the rest of the company to an extent. and I was a part of this early smallish team, which we were really fortunate, like the people who hired me there, like the, now at So some people from those team from that day are now at Parallel. I think it was like a, it felt like a small elite team on a zero to one journey. And so it's not just like working at Twitter, it's working on a problem which was fast moving. That team grew like you would think a rapidly scaling startup grows, right? Like from 10-odd people to hundreds of people in like a year and a half, from running a system on like 4 computers to 4,000 computers in a year and a half, to take revenue from almost nothing to like a couple billion in like 2, 3 years. so that journey was particularly rewarding, that early couple of years, because you get to figure out, like, I didn't know how to hire people, right? Like, but I did so many interviews.

Nakul Mandan:

Yeah, so actually one of the questions I had for you was who taught you to go from an IC to be a leader, hiring, culture, like we'll go into a lot of your operating principles later, but how were those years of learning? Were you bumping into a lot of mistakes and then just learning by yourself? Was there an influence on you?

Parag Agrawal:

Of course, right? Like I think all the— everyone around you influences you. Every problem that you have to solve influences you. I remember Like it was, it was early enough at Twitter that Twitter was not too organized. So for example, one of the things I distinctly remember was like we needed to hire great people. One fine day I sh— and like in 20— was this '11 or '12? One of those, '11 or '12, towards the end of the year, I was relatively new. I was like the most junior person around. And I had this weird idea that like we should go hire undergrads graduating college in IITs. And there's just like no precedent for it at Twitter. used to do like— Twitter had a university recruiting team that went to 4 universities in the US or something at the time, right? and so I was like, I want to go do that, right? And we didn't have a track record of doing video interviews at the time, which today feels weird, but at the time was just true, right? we didn't have too many ways of doing immigrations for people, or we didn't have too much precedent of that. But like, I literally showed up there and in like 6 hours of interviews with a bunch of people, hired a bunch, many of whom became really influential engineers over time. but it was pure people showing initiative and people just going with it to letting you do things that you thought solved the problem of the day, which was hiring amazing people. ASAP who were extremely talented. And so in the process though, for hiring those, I don't know, we hired maybe 5 or 6 people, which was like hiring 5 people in a day for a small team was a big deal and 5 amazing people in a day. But we learned how to do video interviews. I was like the recruiting coordinator. I had printed offer letters for everyone because we didn't know who we would hire. And there was some rule that you had to give the offers that same day, but you could do things, right? You could just go look at a problem and be like, I think this is what I wanna do. And you'd be like, let's go. And so that environment is the one that I think you learn in. I think people think that you learn by only making mistakes. I don't think that's true. People think that you learn by only imitating others. I think that's not true. Of course you have to learn from those things, right? Like you can't repeat every mistake, but you also can't entirely imitate others. But I think if you're given— there's a lot of like, if you decide to solve something a certain way, it is often more possible than you believe if you just keep following through on it, right? If you decide like to do things you can make them work.

Nakul Mandan:

Yeah, but clearly, you know, that environment was there for everyone to do things. So clearly you still rose through the ranks. what was— when we were talking earlier, you'd mentioned like you were just— you're used to this idea of attacking the problems that you think are hard. What's the reason you were able to rise to the first distinguished engineer, the, CTO, the CEO?

Parag Agrawal:

No, I think so one prerequisite, as I've been saying, right, like In many environments, no matter what you do, that is not a possibility, right? And so I think in some sense, Twitter of that era was special in that it was a possibility. Number 2, I think I got lucky to an extent. Like, you need luck for journeys like this, always. The thing I— if I look back at like what I did right and like the common element there, I think it was— it— Two things. One, I think I always like to be with and around really capable people who push me to do more and push— collectively you raise your ambition as well as capability simultaneously. Like, who are those people with whom that happens? And I don't think I would have described it that way back then, but to me it's like a feeling that you look for in surrounding yourself with people and problems of that nature. The second thing I think I was both good at and got lucky with is I think I would look at Twitter and what we were doing a little bit from the outside in. I think today I describe it again, I wouldn't have described it this way back then. But like I wanted to work on, like, you know how every employee at every company sits around thinking like, if I was the CEO, we would be doing X, right? And I had that vibe about me saying like, so what do I think? I would try to figure out how to collect the ba— like how to drive influence, collect the engineers, the PM to be in the place. Where you could, as a tiny group, do things that you believed should be happening or would be happening if you were the CEO, but weren't fully happening or weren't getting the attention they deserved.

Nakul Mandan:

When you were not yet CEO, what's an example of an initiative you were like, guys, come on, we need to be doing this, and then you just did it?

Parag Agrawal:

I think these were simple things. Like, these are not like some crazy big deep insights, right? Like, I used to work on ads. was this separate siloed team, slightly siloed, not fully siloed. And you build ads, you collect a bunch of— you build a bunch of interesting machine learning system because what is ads? It's like matchmaking with machine learning. Twitter at the time used to do a reverse chronological timeline and there was almost a religion at the company to not use too much machine learning when it came to the core product. Right now. So we— I'm there sitting around building ads. I am— I was like a data geek, right? So I am looking at Twitter's going public. We've built ads. Now you can see if you do the math on the growth engine of the company, it to me appeared that it could stall out at under 100 million DAU. The growth engine of the business. So we spent years dealing with, oh, Twitter's growing so fast, we can't roll servers in or keep the site up. And all of a sudden we were going to end up in a world where, oh God, the product is now flatlining in its growth potential, right? And it didn't get there, but you could forecast that in a year that's where we would be if you looked at the data. Carefully enough, right? So I'm like, okay, no one seems to be paying attention to this. There's religion against machine learning. I'm going to abandon ads. I'm going to figure out how to show up on the product side and work with a collection of people where we would go transform the product to drive long-term durable growth. It's like very simple. It wasn't like the, top-down strategic direction for the company yet, but a year later, that's what it sort of became, right? So it's like these kind of obvious-ish ideas, but the right people, focused execution with a small group to prove and change the rest of the company around it. Right. So I did that among— with other people and showed a lot of initiative there. Fought, as you can imagine, some internal battles because these cultural things are ingrained. and I think there were a few more of these kinds of things.

Nakul Mandan:

What did you change in the first 3 to 6 months once you became the CEO then?

Parag Agrawal:

plenty. Some popular, some not popular. The biggest picture, right? When you— I had worked at the company for almost, I would say, 10 years by then, right? Or more than 10 years, 10 years. I always saw my job as looking at the, at the people around me, the problems around me, and figuring out how, I could maximally contribute, right? So there were periods where I would be a solo IC. That was my preference. Even while in leadership, there were periods where I would manage reasonable size organizations, trying to figure out how to hire somebody better to do that so I could go do something else. And so, but I mostly wanted to be flexible in adapting myself to the best of my ability. To maximally contribute to the company or the team. The phase shift you have to go through when you become the CEO is, at least for me, the advice I give and like what I internalized for myself was that you no longer adapt— try to adapt yourself. You challenge yourself, but you adapt the company to who you are, because it is— if you are trying to fit into somebody else's shoes, you're not going to create the best version of the company. And that is a, a, a bad thing to do. You're doing injustice to the potential of the company.

Nakul Mandan:

When you say that, actually, this is a great point you've made. Is that a personality thing? Like, what are you trying to shape the company towards? Your personality? Towards your vision? What? All of it?

Parag Agrawal:

All of it, right? I, I think it's like your personality, your vision, how you like, how you, your mental model of the company. Are you a product-first person? Are you an eng-first person? Are you a data person? Are you bean counter? The three types of CEOs should organize their teams differently. They should hire differently. They should have different kinds of executive team, executive team meetings. Who they hire is a function of them. It should be a compliment to you. Previously, I was trying to shapeshift myself to be most incrementally valuable around the CEO or the rest of the leadership team or the needs of the business, right? Now, all of a sudden, if you take a 5-year view, you have to transform the business. The, Twitter job was— Twitter was not a done deal. Like, there was like a real business to build and scale, and the only way that would happen is for me to transform it into what I thought it would be, right? So you have to own the thing and own yourself.

Nakul Mandan:

What was the thing you thought it needed to be like? Yeah, what was that?

Parag Agrawal:

I think Twitter was already— and I think it's proven today that the product had durability. The product needed to scale to more people. And as importantly, the business needed to get stronger, right? So I think I thought Twitter was too large. So we were actually working on making it a better business over time, partly through being a leaner, more agile team. But partly I think there was so much more machine learning, as I would've said it during that time, to be done. On both the product, on all the issues around content moderation, on ads, I don't think Twitter was like as machine learning filled even for that day as it could have and should have been. And so I think that was part of the hypothesis in, building.

Nakul Mandan:

Were there cultural changes also that you felt strongly about, or did Twitter have a through line in its culture from the days you joined to what it became?

Parag Agrawal:

No, culture evolved quite a lot over the years I was there. I think, yeah, I think I, for example, like you asked me earlier, what was the, the things you did in the first 3 to 6 months, right? To shape the company around me, I made a pretty significant set of changes to the leadership team around me. A lot of it was attributable to me being a different person. Needing a different structure, right? Some of it was attributable to me wanting to transform culture. For example, one of the things I did, I think I took the job end of November or something. In December, I asked the team to do, after several years of Twitter not doing it, some amount of employee performance conversations, management ratings, whatever you might describe it, right? Now these things feel extremely corporate. Right? But I think they are a signal on what you value. And it was important to me in part because I wanted to— Twitter had been a place which hadn't for the last several years done serious performance management, which goes towards both like letting go of people not performing. And empowering and rewarding the top performers. I think there are these two different dimensions. I grew up in an era early on at Twitter where I saw both of these things happen. I was at Twitter when we did a round of layoffs, in fact, two rounds of layoffs, right? I experienced what it felt like going in. I experienced what the company could do after, And those things sort of shaped my view. So I was wanting to drive a culture towards performance. And I think all of these performance and ratings processes and systems at scale are extraordinarily flawed. But the one thing they do is they send a signal. Now, all the data that comes out of it is noisy and you have to exercise great amounts of judgment as you, as you use it. But it's really important to change things. And so month 1, I decided to do something which is not popular at the time, right? Like not among people, not among managers, not among leadership, right? But it was popular among some subset of them and I wanted to empower that subset.

Nakul Mandan:

And then towards the end of that stint, maybe the last 6 to 12 months, you also had a pretty— I mean, in some ways, like few operators have felt that kind of pressure, right? Running the public town square of the world, Elon's takeover bid, the ups and downs of it. I'm less interested in the news story there, the sensationalism, more like the operator, like how was it to lead through that phase of 6 to 12 months when the takeover is happening, not happening, Elon is tweeting all about it, how did you manage to lead the team and how, what was going through your own inner game as a CEO then?

Parag Agrawal:

So I actually think the real interesting operating period was perhaps prior to signing the merger agreement, right? Through that period, you are— we were operating on what my plan would be, which was in fact planning a— I think this is like data that's public and it's out there on the proxy statements that We were planning a significant, reduction in force, layoff at the company. In parallel, there was a public conversation brewing around potential takeovers in the media, which was a little bit off reality, and then a private conversation brewing with the other side around whether that was likely and whether that was real, right? And so that few weeks of navigating these multiple things and simultaneously operating in an intense period. Like the, other thing that was happening then was like the war in Ukraine had just started, right? Which was causing the entire business to evolve, right? So you're navigating the, that reality, the business implications, the layoff plans, and a potential two-sided M&A conversation all at the same time. In a way, all of these were— some of these were my proactive plans, but then there was this confluence of multiple things coming at us all at the same time, right? So that one month was perhaps the, the most interesting or intense. You, in that period, you do learn a lot. Like, I learned about public company M&A, and all of the, things around it in a way that I wouldn't have known. but I think once the merger agreement was signed, it, it was perhaps not interesting. I think there again, right, it takes a period to close. You're kind of in— we found ourselves in a little bit of no man's land. It's no man's land because can't paint the long-term vision, especially in a case where there isn't deep cooperation on the two sides, which fell apart within a week, right? So now, the best way to run a company is to have a long-term plan and an ambition and a vision and a collection of steps that get you there. That put you on that path to learning. Like, that's the only thing I know, right? And once you take away your capability to credibly describe a long-term ambition, vision, future for the company, that's a hard operating environment, right? and you had lawyers on almost every decision, knowing that this was looking like a contested moment. Now, credit to the leadership team there, like Twitter continued operating well. We managed despite the uncertainty, every advertiser's feeling the uncertainty and wondering if they should pull out. Our teams managed to not let that happen through this period of uncertainty. We shipped a lot of valuable things in that period that, I'm proud of. We managed to do a lot of performance management in letting people who were not performing go through that period too, rigorous— through rigorous management. So because like the, framework is like the deal isn't closed until it's closed. And so you are now operating in this weird dual reality of— listen, like, we are committed to closing this deal and we'll put in every effort in the court of law to make sure it closes. And simultaneously, we are operating like it won't, because one day we might be on our own and we have to have momentum, we have to have great people, we have to be shipping things, we have to have product velocity. And so you're now— and I could keep these two contradictory things in my mind. It is very hard to have thousands of people operate in this sort of weird dual reality.

Nakul Mandan:

So how are you communicating that? Because I agree that dual reality seems like extremely tough, and it's kind of, also maybe not a replicable thing. I don't know if there are any lessons that you can take to parallel from that.

Parag Agrawal:

No, I think there are, there are base level lessons you can take, right? I think I generally was, to the extent possible, I was, honest and direct with people. You can actually go see what I was doing because every time I did a lot of all hands through this period where the entire company got to ask me questions. I did several of them. and they all leaked, right? Like there were, there was media in every one of them. And so, so yeah, that's out there, right? It's just a weird environment. I think the lessons you take away from that are sometimes like you, you don't have agency on certain things like that, like fiduciary responsibility takes over. Like it's like, To, to, be only a fiduciary, right? But I think at some point that's what that job became. When I took the job, I didn't, I didn't think it was likely that that's what the job would become, right? And I, if you told me that's what I was signing up for, it was like a primarily fiduciary for a period, like that's like not the most exciting thing when you're trying to build for a positive sum game to end up in this sort of zero-sum contested M&A transaction, even if that's the best return in the board's view for the shareholders. yeah, so I think it was not necessarily something you can take a lot of lessons from. Doing right by people, like all the people that I worked with, whether or not we made hard decisions with them, I feel pretty good that today if you go talk to them, people who are at least close to me and even a few steps removed from me think I conducted myself in a way that I'm proud of.

Nakul Mandan:

So you told me earlier you had never planned to be a founder, right? Till that moment. So what post-Twitter, what went through your mind? I think you went quiet for a couple of years almost. what went through your mind? Why be a founder at the age of 40?

Parag Agrawal:

40 is young. why be a founder? I, I don't know. So a couple of things. I am— when I took the CEO job at Twitter, it was like a, like a real process, right? Like you have a bunch of interviews with the board. You have to think about what you're going to build, what your vision is, what your ambition is. It is like real intense 6 to 8 weeks prior to announcing of intense work. And in that work, what at least my process was like, I put a lot of fuel into myself and a lot of energy. And like, you like load yourself up with like energy to go for— before you get to refuel again, it's going to be like 5 to 7 years and you're going to go all in and do everything you have because that's what it'll take to transform this already large consequential thing into something it could be or should be, right? So now you have all of a sudden all of this unspent fuel sitting around without a clear output channel for it. Like, which way do I point this thing, right? Like, I put all of it in over this intense period. Now I, I have all of this reserve. Like, what do I do with it, right? And I was pretty open-minded, like, initially, but I was, like, hacking and building, and it's like I've said before, it's like a good time to be a technical person who was unemployed. It was, like, as good a year as you— one you could choose to be in that position, right? and in that position, it felt like the world was about to change and transform. And whenever I looked at working in places, it felt to me that one, I wanted to work with like a small number of people, like this whole desire of like special small number of people that I get to choose. That is almost the most important criteria. Two, something technically hard. Where my— I wanted to be much more technical again. Like, I think you lose a little bit of touch, even if you try really hard, to the technical work. If you are in a large place as a senior executive, you can allocate only so much of your time to it. And I really wanted— like, I was starting to write a bunch of code again. I was really enjoying it. Being in the weeds on like just technically solving problems with my own hands, in like a real way. and so like all the big jobs felt to me like they didn't like, they didn't pull me in. And so I relatively quickly realized in like maybe 2, 3 months that like I would have to become a founder. But then it was not clear what I wanted to do.

Nakul Mandan:

Well, what convinced you on this idea?

Parag Agrawal:

I think the idea found me. I was perhaps like, I wasn't proactively looking for this idea. I was proactively perhaps looking for other ideas running in other dimensions. And this was like the side hustle, which was just like my hobby or fun, like write fun code. And then, you know, you're trying to understand various markets and dynamics and finding a point of view on various things. And, you know, I'm the engineer who ended up being like, okay, I'm going to actually write an agent to make my life in doing so easier. And at the time, agents didn't work. We were on pretty bad models. but I found myself spending more time doing that rather than actually doing the work I signed myself up to do. Right. And then over time, this thing like became a— obsession of sorts. And then you start thinking about like, oh, these are all the problems I see, but really what will happen if agents work? They don't work yet. What does it take to make them work? If they work, what happens? What happens to the web if they work? And it started to germinate that like agents will eventually work. I had spent a lot of time with like labs and researchers through this journey that like I was a believer that models will keep getting better, agents will work. If they worked, I made up that they'd use the web 1,000x more. And if that happened, you'd have to invent new technology. Really exciting. New business models. Really exciting. And it felt like I had extraordinary confidence in that moment that this was just going to happen. And it was like, to me, it felt in that moment obvious. When I spoke to others, it didn't yet feel obvious. So the same trick of, oh, Twitter needs machine learning in the consumer product, because like, how else do you use people's time most effectively when they choose to spend 5 minutes on Twitter and show them the most relevant fun stuff? It felt obvious it wasn't happening. This thing felt similarly obvious, yet others around me didn't believe me by and large.

Nakul Mandan:

So let's double-click on it. This is Parallel web is, search for agents, right? In a— why do agents need a different web search? What's different about how agents search than humans?

Parag Agrawal:

A lot, right? First, let's take the top-down perspective. If agents are going to use the web 1000x more, can you build 1000x the search infra that Google has built to support it? I bet no. Okay. So now you, you have to have an argument that you can do something novel and unique. Okay. So you ask the question, what will this thing look like? And here's what you have. You are seeing the best capability to show up at the hardware layer in GPUs. You want a lot of memory, you want lots of memory bandwidth, you want flops, get GPUs. What if you decided to frame the web search problem as a problem to go solve it primarily using GPUs over the generations of commodity GPUs? Second, these models are getting better. The other thing that's happening is holding capability constant, you're going to be able to get— make the model smaller and smaller and smaller. And so now you will be able to distill into a much smaller model, a lot of intelligence in a very broad-based way to build your anchor. So this is the dimension which is new hardware and technology advancements as it relates to AI is going to allow us to build something which you couldn't have built a year ago or 2 years ago. That's number 1. Number 2, agents don't behave like humans on the web. So if you think about it, let's talk from a few dimensions. Inputs. A human going to a search engine, their input is a short, underspecified search query. You're sometimes even using autocomplete, right? So mostly we don't tell the search engine what we want because we are impatient and we haven't even thought about it, right? Like you're trying to do something and you like write two queries and then you go on this random walk. Of clicking on links, reading paragraphs, going back, searching more. That is how humans work. Agents are happy to write, sessions worth as a query, right? For human search engine, the output is 10 blue links so that you can go on your random walk, right? For an agent, the output is a collection of— let's call them tokens to put in the context window. That give the agent its next set of action space, correct? Which isn't just clicking on links, right? Number 3, think of this human search engine. It monetizes at a certain rate. It has a certain latency profile to it. And those 2 things roughly tell you what kind of compute to throw at the problem and what kind of architecture to use, right? So it monetizes better than I think agentic traffic will, because agentic traffic will be 1,000x. Two, the latency profile on humans is— humans are like a page loading between 500 milliseconds. It's great, right? It takes 2 seconds, it's really slow. And you'll be annoyed. If it takes 400 milliseconds instead of 500 milliseconds, you won't even notice, right? So there is a sweet spot you hit in page loads in search. With agents, there's high variance here, right? Like, you think about a voice agent. If you take 500 milliseconds, 600 milliseconds for a voice agent, it feels too long a pause before the agents can talk back to you. If you just do the math. Also, the entire premise of agents is humans essentially have a 1-second SLA in life, right? Like we want things to be interactive on the web, loading a page, talking, voice agent coming back in sub-second. Agent's entire promise is you're going to be able to do more within each turn, right? So you ideally want search in an interactive setting to go from this sort of 500-millisecond thing to a 50, 100, 200 millisecond thing depending on the use case. But in the background agent context, search can take an hour. Who cares? It's working overnight, right? So now you have this extreme variance in what could be useful and this open canvas to build entirely new technology around entirely new optimization points. The input is different, the output is different. Your compute budget is different, your latency budget is different. What we've solved historically is this one narrow slice of it.

Nakul Mandan:

So it does feel, to your point, a lot of moving pieces, but what's the, can you actually talk about like what's the hardest part you've realized in the years of building now, like which you may have even underappreciated when you were getting started on this journey? What's the hardest technical part?

Parag Agrawal:

I think the hard technical path is, so one, right, like you don't, you have a lot of conviction and confidence in agents will use the web at scale. And a lot of it will have, like, you can— a few of those things you can take to the bank. And we had conviction on the details, right? How fast will that happen? Which models? What will— like, I did not have the foresight to know that coding agents would be dominant in the way that they have been for a period, right? Like I would not have anticipated that. So now as you see these things unfold, you have to be agile around it while being opinionated about the long term. The second thing that's really hard is that you have to break— to be in this business, you kind of have to break a bunch of rules, right? People will tell you, so we are building a horizontal web business for a lot of customers that look very different from each other. We are telling the world that if you are buying models or LLMs, using our search to enhance those tokens makes them more capable, whether you're applying them for finance or legal or insurance underwriting or sales or recruiting or science, or you come up with any use case for us, a human in a browser, which is basically everything, right? We are saying we are building a product that is useful across all of these modalities, right? So now these are very different kind of customers, very different kind of sales. Very different kinds of value prop even in terms of the language you need to use. So building a really horizontal technology layer from day zero instead of ending up there by saying, I'm gonna go do a vertical and then gradually expand it is a different beast. And you kind of have to break a lot of well-trodden startup building advice and pathways. To go there and have that ambition.

Nakul Mandan:

And Google has the compute, the talent, the background, the index. Why isn't this Google's war to win by a long distance?

Parag Agrawal:

I think the— I think you should invert the question.

Nakul Mandan:

Okay.

Parag Agrawal:

Like, make— I, I, why would it be Google's war to win? As I said, the problem is entirely different. I think anytime there is such a rapid rate of change, my entire hypothesis as why I wanted to be in a small thing is because a rapid rate of change favors a smaller AGI thing. two, I really think our goal, right, is to allow everyone's agent to be able to have access to the best search and all of the information on the web, right? Fundamentally Google's goal is to allow it only for Google. If you are now running, open model on your hardware and you're running an agent there with your own inference, I want to provide it search. I wanna provide it access to everything that anyone else could. And that is fundamentally not any of the larger, better capitalized people's goals, right? So in some sense, how we build our— product, the pressures we feel, how we get feedback from the market is entirely different from how somebody else does. So then you can ask the question, do I get high fidelity, higher signal feedback? Do I attract better talent as a result of this strategy? There is a right level of, in my opinion, resources and resource constraint. Relative to the problem you wanna solve. And in this case, I get to semi-orchestrate that to create the conditions for the best thing to get built, right? Like, Google doesn't exercise any of these things or control. They're more forced to go with the flow while we get to create the perfect conditions for this to come into being.

Nakul Mandan:

So with Google, on your first point also, the, Google point, I agree. they're big, stodgy. They also have their own, not just bureaucracy, but economy that they've built. But the labs, this is pretty critical infrastructure for the world they are building. You're building— why wouldn't OpenAI, Anthropic want to own this? It's like central. It's the world they're building and they are pretty agile and they are pretty execution-oriented right now.

Parag Agrawal:

So one thing I said about Google still applies to Labs, right? Like when they build it, they build it for their models.

Nakul Mandan:

Yeah. So it's model agnosticity.

Parag Agrawal:

So that's one part of it, right? The other part of it is if you think about what is existential for them, it is the model, the layer above the model. And this is incidental in my opinion. not because it's not important, right? Number 2, the mission to democratize all information, like it doesn't flow from AGI where you're saying your worldview, world model is I will get a model which can do anything a human does and I will make it so agile that it can deal with every bit, every tool that we have built for humans, as perfect or imperfect as that might be. And I will boil everything into this one thing, right? Our world model is these things are going to be amazing and exist, but there are some tools like web search and the web that instead of making the model be good at adapting to this tool, you gotta adapt the tool because it's pervasive. To the model. And I think sometimes, like, the point of view and the worldview collision, if you're right, gives you the edge. So now it boils down to, is our worldview right? And if our worldview is right, we will prevail.

Nakul Mandan:

Your thesis relies on the internet remaining an open place, like, but there are big giants are trying to be more walled gardens, right? Recently Amazon with you know, Muse and Instinct is trying to be a walled garden. LinkedIn has always wanted to be a walled garden. we do a lot of recruiting here, so we know how much of a walled garden— we actually use Parallel Web, so we are a customer. So do you feel if the large internet giants end up being walled gardens, that's a problem for Parallel? Or do you think some— at some point they are forced to flip? There's no way they can remain walled gardens in the world of agents.

Parag Agrawal:

So actually, I think it's perhaps even more interesting to talk about, like, almost everyone on the web and their incentives. So there is a real risk, right? Like, the web, the open web, including many of these large players, are built on a few business models, right? Like, what are the business models on the web? You have ads, you have subscriptions by the seat per human, you have e-commerce and transactions, which could be for digital goods or physical goods, right? And then around transactions, you have like a mini affiliate economy or a payment economy or a, economy around sort of the payment flow, right? Like these are the core dominant business models. Now ads and subscriptions per human, it's very hard to imagine those surviving in their current form. Into a world that I imagine where agents are the primary consumers. Now, e-commerce transactions, payment rails, perhaps even some affiliate stuff could survive and should— like, I, I imagine them surviving, right? Even there, there's a real problem around broad-based matchmaking because previously you could have tolerated limiting your universe of what you would matchmake to based on committing to a merchant or something. But I think with the agents, that thing is going to get so efficient that you will exercise matchmaking and selection over the entire universe of availability. and so now we have to figure out how do we keep the web open? Because if we don't quickly build a incentive for people to have their content available to agents, Invariably it's going to shut down. You're like, oh, agents are stealing my stuff. I used to get paid by ads. I used to have people click through and buy subscriptions. Since people started using agents, anytime I put something out there for free, people steal it. I get nothing in return. Why should I keep doing this? Right? That's going to— and it's a very natural question, right? and I think that's part of what, like, I think the mission of the company is to keep the web open for agents, for everyone's agents, and not just And then, you know, you shut it down for everyone and then you go to OpenAI or you go to Google or you go to some other large player and you're like, okay, I'll give it to your agents if you pay me a bunch of money. And they do. And some content owners for a moment get paid, but then they stress about like, okay, in 2 years, this business is— the agent business is going to 50x, 30x, whatever. They're not gonna pay me 30x.

Nakul Mandan:

But what's the incentive for them? I agree with everything you just said, but then there's a huge misin— disincentive for LinkedIn, Amazon to go open, right?

Parag Agrawal:

Yeah. And so, so how will that work out? We have built this thing called Index where we have built a model to price every bit of content that an agent reads. In terms of how much it helps the output. And I think that is now an interesting currency in town, right? You can usually for most places, you can say, okay, if, I spent a little bit more compute on the problem, I could have gotten a marginally better answer. That's something we sort of intuitively now see in the data. We also believe it, right? Like, okay, if I use a Opus versus a Sonnet or a Haiku or an aphra versus Sol or Luna, I trust the results of the higher compute thing more because they empirically tend to get more things right and do them better. Right? So now you, you're living in this world where when you do work and you produce output, you are saying, if I spent a little bit more money on this, I could have gotten a marginally better output. Now I can trade that saying, okay, if I had access to your data as I did this work, and it made my output marginally better or more desirable. Oh, I could have done that with a dollar of compute. So you added a dollar of marginal value. Why would I spend compute when I could have paid you 75 cents on the dollar to get that data, not used up the GPUs? And so there is a bunch of data we've collected around this. We've trained models around this. And we want to ultimately incentivize the creation of high-quality content because models are not very useful if not given data and information, right? So you want them to have data, you want the people to have incentive to create the right and high-quality data and make it accessible to everyone's agent. And you wanna do it at scale.

Nakul Mandan:

So this is almost like a new business, like advertising business model for the new internet.

Parag Agrawal:

It's, a version of, you can say, aspirationally AdSense in its early days, is what we think we can and should build here because otherwise, like, you're going to have to pay some large lab or something rent to access this data. And in the process, you lose and the content owner loses.

Nakul Mandan:

What's been the reaction of the big guys, to this, the walled gardens, I guess?

Parag Agrawal:

I think it's, in flight. We have— we've announced a collection of partnerships. We've not announced any partnerships with like a large walled garden as you describe it. but I think they feel and see the problem too. we've announced partnerships with some amazing partners. Like I think we have, we're working with like The Atlantic and Fortune. If you think of like news publishers who have built real businesses over, I don't know how long, we work with really interesting data providers to different verticals. for like, you're a VC, so you'll understand we work with PitchBook and Traction. these are datasets which would otherwise be, like subscription by the seat. and we work with them to figure out like when their data with their authority and quality produces marginally improved quality, how do we make sure we pay them? and increasingly we are expanding to now lot, more of these content partners. But our ultimate goal is to have a super scalable, business model that not just makes these businesses survive, but thrive with real participation in the agent economy. And again, that only happens if the first premise I've started with, which is agents will use the web 1000x more. So when you grow the pie a lot, you can use incentive alignment mechanics like the ones I'm proposing to actually incentivize everyone to collaborate.

Nakul Mandan:

I'm curious for your take on 3, 4 years out, to the extent we can predict anymore. it's a crazy time to make any kind of predictions. What does the internet look like, or what does our world look like in 3 years? Do we only talk to agents, never go to, you know, a website really, other than maybe social media like Instagram?

Parag Agrawal:

Yeah, I think we spend most of our interactions with the web as we think of it today happen via agents. I think we continue to use experiences online. Those experiences perhaps don't really look like a traditional HTML website loading. They perhaps look a little bit more like an app, perhaps with more AI blended into it. But I think there is the— it would be rare-ish to go to the web to do work, right? So you'd go to the web for what I'd call experience or connection, some of these other needs. I think the biggest thing that to me feels again, totally obvious about the future 3, 4 years out is the predominant ways agents interact with the web won't just be the, the search or the pull as I call it, like agents pulling information from the web. And will transform into a push where your web system like Parallel will have enough context to know what you are interested in. You've kind of subscribed to a feed or a push stream, which tells you, oh, this is what just changed in the world, which now requires your agent to do some computation, to bring something to you, the person, or you, the company, right? So this transformation of the web from pull to push, I think is what we've started seeing. And I think the, the first, like until what, like few months ago, there were no agents which felt like they were always on and proactive. Now we are starting to see the very first set of agents which are creating this abstraction for you or an expectation for you that they are always on, they're proactive, and they're trying to know from your data and learn from their interactions with you what you might care about and go bring those to you proactively. Right. I think if this plays out in the way I imagine, The web has to change to go from this sort of pull to a push model. and that I think is like what I'm really excited about. Like that is the singular thing that we think we will end up building.

Nakul Mandan:

Do you think Google is toast in this new world?

Parag Agrawal:

So again, right, like I think the, word toast is loaded. I think in this new world we are talking about such— a potential expansion of the pie that you can simultaneously keep growing while losing share of market. In fact, like share of market is like a vague thing to define because the market itself is changing, right? So yeah, can Google— if you think of like today, The web search market, Google commands a hefty market share. Should and can Google command that market share of a future AI web search market? I would like to think no, obviously, but can they still grow their business even as this new market comes alive with its own set of Events. Yes. So I think you can, you can totally imagine a world where Google with their accumulated capital, talent, technology, culture, momentum, distribution continues to build an increasingly compelling business while new people like us create entirely new markets where we get to, participate and play. And what we are betting on is like over time our market is bigger than their market.

Nakul Mandan:

Yeah. Might have to go through a Microsoft-style transformation where, you know, after— under Satya, Microsoft has gone through a transformation of product to services and GitHub and many other things drive it versus—

Parag Agrawal:

I think this world is moving so fast that almost every business will have to undergo transformation. In fact, like between now and 4 or 5 years from now, I might have to go through two transformations myself, right? Like even as the new kid on the block with the new potentially agile business, like we like to think way more agile than anyone else we're talking about. In this market entirely premised on this wave, we will still have to go through transformations.

Nakul Mandan:

I want to shift gears towards your operating principles. first of all, how has being a startup CEO been similar and different to being Twitter CEO, public company, a large company? Maybe start with similarities.

Parag Agrawal:

Today Parallel is getting bigger. What,

Nakul Mandan:

around 150 people? How many?

Parag Agrawal:

Less, 120 maybe. and it, it is post product market fit on part of our business, right? So I think in the early days of Parallel, it was very, different. I think now you're starting to see some of the elements show up. but I think the common themes, right, are ultimately most leadership. Is about having followership, right? Like, no matter what job you're in, if you're in what would be considered a leadership job or role, you're not going to be very effective unless you have your own brand or way of getting followership from people who can collectively do something special. That is at its premise common, no matter what you are. So now the question is, do you need— can the same person be both? And I think we've seen time and again that even if the things that matter might be different, but most entrepreneurs who end up scaling or doing multiple things prove that in different circumstances, there is, there are some common elements. And people have many different— there isn't one brand of leadership in the world. Each person builds their own way to find people, and people who end up finding the right people to collaborate with them in the right time end up doing interesting things. So those things are the same, right? I would say almost everything else is perhaps different, right? Like post-product market fit, pre-product market fit. Everyone cares what you do, no one cares what you do, right? Entirely technical hypothesis, there is no market, no customer. Like, what do you do? How you recruit? Who you recruit? Everything is different.

Nakul Mandan:

What have you had to unlearn in your own behavioral styles as a leader? To recruit, behave with your team in a technical problem versus a post-product market? Yeah. What's changed for you as a CEO?

Parag Agrawal:

I think, I think the early days of Parallel were perhaps more natural for me, right? Like as the technical person among technical people. Smart people, some deep conviction in something world doesn't believe, building something that felt natural in the zone. You don't like— so you don't think about things, learning or unlearning, right? You think about learning things. But in that sense, those activities felt easy, natural. Like I was sitting there like, how do I unlearn things? Right? What I do think is a common theme for how I work is I like doing— I think I'm only good and I only like jobs where I think that I am slightly unqualified for doing that, right? Like when I took a software engineering job at Twitter after doing a PhD, because I, I felt like I was underqualified in that moment because like I had done like academic research, I had done prototypes, like real software engineering for real systems with real great engineers was a real— it's a challenge. As I took on like every job at Twitter, I felt that someone was taking a bet on me in a way that like me in my most rational moment should not, right? And so you only want jobs, in my opinion, that you feel slightly unqualified for, because that drives, at least for me, all of the right behaviors, right? Like, you are curious, you work that extra bit harder because you're not up to it. You aren't rinsing and repeating a thing. You are there with a curious open mind about like, how do I go tackle this thing? You seek help from other people or from models, right? And so to me, that is how I felt in every job at Twitter over time. Being a founder felt like that to me, right? Like I'm founding a thing, I'm—

Nakul Mandan:

what are you unqualified for now?

Parag Agrawal:

That is my current job.

Nakul Mandan:

Yeah,

Parag Agrawal:

entirely unqualified for like scaling the company. I'm like entirely unqualified for scaling, API business for developers and enterprises and growing at rapid scale, a large team in a rapidly evolving world. Like, I think like anyone who thinks they have been trained for this hasn't been, right? Like, I think these are— fortunate enough that we found ourselves in this unique moment, in this unique position. To, go on this ride, but like I'm unqualified for almost everything I am supposed to be doing.

Nakul Mandan:

As a CEO, are you spending your time these— like, how does your week actually look like? Do you have structure around it? Are you one of those who attacks different problems every week and there's not as much structure? Do you have weekly meetings?

Parag Agrawal:

It, has evolved a lot. Things were very different about 6 months or a year ago. When the team was materially smaller. Like our team was grown 4x this year, I want to think, in like 9-10 months. So it's, a work in progress. The evolution of me, I do— the core structure of the week that I care about the most is I like seeing engineering demos on Friday evenings. I don't know, it's, it makes the, like, I like to see what is being built or what got built or what got shipped to connect it to some customer value. I, that I really care about. I spend time on Mondays often with my leadership team, mostly because Every weekend is a new obsession or journey for me to— like Monday to Friday, I think you're in the operational day-to-day and every weekend you take like one step up and come up with what are we missing? What should I be paranoid about? Should we be going and creating a new area of the business? Do I need to go hire somebody with a special skill? Do I need to think about a customer differently or market differently? So every weekend you end up in— at least my usual goal is to select one obsession from three that are bubbling and try to come up with a conclusion by Monday morning, which is either this obsession is not something I should act on, which is the right answer half of the time. Or this obsession seems worth pursuing, that I am going to hand something off to somebody to take it to the next level, and then I will figure out some way of following up on it, right? So To me, that is the— my cycle is a little bit around the— this weekend and the rest of the week. We have a, we have a company thing where we are in person in the office, in a single office, and everyone is in the same office in product and engineering on Mondays, Wednesdays, and Fridays in Palo Alto. People who live in San Francisco go to our San Francisco office on Tuesdays and Thursdays. Right. So now my internal time with people, I try to reserve as much as Monday, Wednesday, Friday as possible for that. External stuff to the extent possible, I like to do on Tuesdays and Thursdays. So there's some natural cadence that gets built as a result, right? generally I don't like scheduled— almost no scheduled one-on-ones.

Nakul Mandan:

No weekly one-on-ones.

Parag Agrawal:

No one-on-ones with anybody. We're pretty flat. no one-on-ones. We don't have people that need too many one-on-ones, but It's always one-on-ones, right? Like I'm always in the office and anyone walks up and you can have a 3-minute conversation about anything. And I mostly process things by— as I'm walking around the office, I see a face and then I'm reminded of like, oh, there was something that was brewing. Let me go figure this out with this person.

Nakul Mandan:

How much of your time are you spending on recruiting now?

Parag Agrawal:

Recruiting is always a steady amount of time, but then there are periods of intense activity.

Nakul Mandan:

But is it like 25% or—

Parag Agrawal:

In aggregate, yeah. But I think there are some weeks which will be lower, but certain weeks where there is like one job, right? Like I think we've recently, I think now we in the spring got our amazing head of sales, Graham. Most recently we got our head of marketing. Elon, right around then we got head of finance, Odil, really amazing people, but each of these are consequential hires. So each of these you obsess about a lot of heartburn about like, is this the right person? Is this the right thing? a lot of time with the person, but also a lot of time making sure you know what you want. Are you going to set them up right? So recruiting isn't just about. The interview, right? Like that is, I think, like a sliver, right? Real mental energy and time goes into all the other things and half of it is done after they show up.

Nakul Mandan:

You talked about hiring for two types of alpha when we were chatting, before this. So can you double-click on that? Like what's your philosophy on what you look for?

Parag Agrawal:

Yeah, so listen, this is like, goes back to Twitter days, at least my biases. Is I always thought of— listen, like when you are the person that wants to get hired to do jobs that you're not entirely deserving of, and if you decide that for yourself, that is the best way to maximize your career, your prospects, your impact, whatever you might want to do it. Then you have to simultaneously believe that you only want— you want to be able to hire those people, right? Like if there is someone, I have to figure out the best people I believe will want to be hired for things that are a challenge. If they have been there, done that, do they want to do that thing again? Or do they want a new challenge, which will push them on a journey to learn and grow and do something that will give you the satisfaction saying like, oh, going into this, I didn't know that I could do this or we could do this. Doesn't it feel great when you do it?

Nakul Mandan:

So you look for more potential rather than they've done the job.

Parag Agrawal:

Correct. So that is what you mean by alpha, right? Like you have to make bets on people which feel like bets. And, but now if you're in the business of hiring a lot of people where you're just making bets, right, you are mostly hiring some form of outliers. Now you also have to pull the team together, a team of outliers together, right? And so now you have to ask the question, who are the people who are exceptional at building these unique weird teams where you have bet on potential. And there I at least find that I want to see on that dimension evidence that they have done that. They have worked with teams of exceptional outlier people who aren't of the type of been there, done that, and created an environment where they can all do their best work and grow and learn, which is not like default, like management best practices, but like creating sort of these special teams that create this special moments, right? So you hire people who've done that, but then you challenge them by my principle on a different dimension, which they might not have encountered, right? So that's sort of how you build teams. But like, I think it's like a, It's fundamentally a slightly non-scalable way of thinking about team building where you're trying to take bespoke individuals and creating something out of that instead of saying like, oh, I have this thing, I'm going to cut it up, divide up roles and responsibilities, which requires a different kind of agility in leaders or managers, different kind of agility and plasticity among the people who show up and do this. But I think that's what the— early special teams are built off.

Nakul Mandan:

Yeah. But do you then rely on interviews? Do you rely on backchannel references? How, do you figure out that alpha that you're thinking about this potential is actually there? Did they actually build those weird, exceptionally talented teams? Were they able to manage it?

Parag Agrawal:

I think you— we do interviews, we do backchannels. I am— okay at it, I would say, at the hiring moment.

Nakul Mandan:

On interviewing or backchannels or the overall thing?

Parag Agrawal:

The decision, right? Like, should I hire this extremely high potential engineer? Let's say that's the question. I think when I go and make that decision, I am okay at it. I think our head of engineering is materially better at it. I am perhaps better. At leadership hires. And so I think— but what you have to be, if you're going down this path of betting on a lot of potential and moving fast, you have to be willing to admit mistakes and failures. And so I think I am good at admitting to mistakes and, like letting go of people. I think like they're fundamentally need fixing and to keep a really amazing team, you have to be willing to admit when you've made a mistake and letting people go.

Nakul Mandan:

But on the exec hiring, you said you're better. Why, are you better? What gives you that edge? Are you looking for something that you've learned to look for better than—

Parag Agrawal:

Yeah, I think there is— I think over time I have, at least my interview experience, perhaps, biased towards exec hiring. So I feel more on my feet and grounded. I, I think you fundamentally, while ex— hiring execs, right, there are two things that happen. One, there is more raw data and evidence relative available in the world about what journey they have been on, which you can elicit from them or from back channels or others they've worked with. So one good thing is you have more data available. Two, you have to look past the— when, like, the outside, whether it be in an interview or in a back channel, to get to the data, right? There's this, like, enough polish, enough experience that you need to look past it. I think I've gotten somewhat okay. Or better than many at doing those two things right on exec hiring. So when there is a data-rich environment, I think I can look past the, polish to the raw data and make some extrapolations from there, which I feel— I feel grounded on my feet and I feel good and conviction when, we hire people.

Nakul Mandan:

On exec management, then how do you— what are the earliest signals, you know, that this— I made a mistake? Because you, your philosophy includes correcting mistakes?

Parag Agrawal:

I think it's not hard. Like, I, I actually think getting hiring right is extraordinarily hard, right? Once someone is around you working, we currently work in person, we work with other people, like, there can be some gray area, right? But the worst mistakes are very obvious, and they're not just obvious to you, they're obvious to everyone around you. And if you've created an environment where you can walk up to an engineer and ask them their opinion, and if you can trust that they'll tell you honestly— again, to my point of eliciting information and creating an environment where I can get to the truth— it's not hard. It's just not hard. And so it becomes harder if you create an environment where no one tells you what they think. Where no one even acknowledges what they believe, right? But if you don't create that environment, it's actually very easy. So I don't know, this is just like, for us, it just doesn't feel like a hard problem.

Nakul Mandan:

How has your role evolved in this AI era? Like, let's talk about AI-native organizations, AI-native execs. You were a CEO before AI. How are you using AI for your own life as a CIO, that is?

Parag Agrawal:

Yeah, it feels amazing, right? Like, so one, I have over this period sometimes run ahead of AI's capability and for some weekends been in AI psychosis, like many of us, right? And then some, like reality brings you back. Saying like, oh, I thought it worked, it didn't. So that has happened. I do think— I think the biggest superpower for me, the way I would put it is, listen, as a CEO building a company, you are paranoid about everything and you kind of— in an ideal world, right, I want to transmit insight context to my entire team that I have my convictions about the future. There are so many intricate details in your sort of worldview or details or how I want to show up to every customer interaction or every principle in our API design or technical architecture design. I have somewhat strong opinions on all of these dimensions, which are kind of upfront, like you can't write it in a book. And hand it to people and transmit it. It's just like, I don't know how to do that. You provide higher-level frameworks, you provide information context, but the only way to reinforce those things is to observe things in a moment and then inject your point of view in the right moment, have it be visible to a lot of people so that people can extrapolate for it— from it. And bring your worldview. Now, what does an AI give me? So what I'm fundamentally describing is you go in, give people a lot of autonomy, but occasionally go in and micromanage on specific things. When you see things, when you either need to reinforce a behavior or you need to micromanage something to tweak and provide more context to find that opportunity and that circumstance. Now, AI gives you extraordinary capability. Like I would not be able to keep up with the code being shipped. Now at some high level, I maintain a mental model of what is shipping by having a lot of AI summarise it. Like even if like 10% of it is misinformation, I have a week-on-week mental model of what we are building across a high-velocity largest team. I have AIs looking at any visible interactions to, like call it a Gong or Slack interactions with our prospective or current customers. I try to pull out insights from them. I observe what is the way I would like things to be. What are things I wanna learn to put on the product roadmap? So it gives me this extraordinary ability. To prioritize my attentions and learn things to know where to go in and tweak something. So I feel like way more control on the thing we're building than I would have in a prior life.

Nakul Mandan:

Do you worry also that it's over-information, overkill of information? Because the other side of it is all of us have so many agents these days. Pushing us more information. Is that happening to you? That, not really?

Parag Agrawal:

No, listen, your time and attention, mental energy is a constraint. So now earlier I was on a— if the prior is right, like you were getting a random subset of information. Now I'm using AI to prioritize it based on some things I can say. So now I spend enough energy to— so that the AI is curating something better than the random walk I would otherwise be on. Right. So, so I do think for me, AI is by and large a really material thing in how I operate or how I spend my time or what I do with it and what I'm able to do at the company, be in more details in much more places, have a— far fewer meetings which are information gathering oriented, have more decision oriented and influence and steering.

Nakul Mandan:

Yeah. What does the role of— how does the role of engineers and engineering leadership evolve in this new world?

Parag Agrawal:

So one, I, I wouldn't— I'm not— I don't feel like I can call myself an engineer. Right. I think it's evolved rapidly and continues to. I think there are certain things on engineering which now AIs are good at. There are certain things where they're a function of how you architect your system so that the AI can be good at it. For example, right, like you would— The biggest change, at least for us, is, you know, you think about making engineers productive. And I think every time you scale an engineering organization— I was the CTO for a while— you think about like, okay, how, how do I make every developer more productive, right? Like, what are the base things to make every developer more productive? And there is some maturity curve on that, Anthropic and organizations have gone through, right? In this world of AI, you kind of have to speedrun that curve, right? Okay, my engineering team of 10, you should actually invest in the developer productivity things that you might build for a team of 300. Again, like there's the real difference is to guardrailing for agents versus humans, but like at scale, at 300 scale, you start building systems where you're like, okay, one engineer will make a, a crazy mistake. Which is roughly an agent will do something crazy or is more likely to do something crazy. And so to me, that is the biggest way, like more of engineering is to be thought of as scaling engineering and building more of the meta stuff.

Nakul Mandan:

Are you seeing people on that curve very rapidly because they have to change their mindset? You know, like in terms of the guardrails, the engineering productivity, and really thinking that this is a team of 300, actually not 10.

Parag Agrawal:

That's something— I think on our team, we have the right mix. We have, we have parts of the stack when you're building a very large-scale crawl and index and these super scaled, complex distributed systems where they're very stateful, every deploy is complex. Like there are these complex pieces, low level, where I think we have failed to make it iterable by 100 people. Like 100 people can't iterate on this complex beast. It's just not— dev environment doesn't exist, right? So there exists a part of our team which is working on those. Another part of team is moving really, fast on iterating on quality. And so within our team, I think you would say there's some variance in use of tokens, use of AI. And I think that's a healthy thing. Me, I'm in this camp of like the token maxer camp, right? and then there is the responsible camp, which is like, oh, I need to build a reliable system in the presence of these token maxers. So I'm gonna go obsess about deploy tests and CI and security and regressions for every customer that exists as the rate of— as the velocity shifts, right? And to go faster, you need to do that work also. So I think we have this team which is doing that. Now, of course, like, I think we could be a— like, every CEO thinks this, like, we could be like 3x better. We could be 3x better than we are, but that is to work for. Like, I, so I don't think like we are at the peak performance yet, but I think we're doing all the right things to approach it.

Nakul Mandan:

Let's talk about the future of the world, right? Like you're at the center of this AI revolution. Are you in the accelerationist camp? Are you in the Dario camp that, hey, we need to slow this down before I'm guessing based on your token maxing statement, you probably are in the Jensen camp, but like, yeah, where are you in this? Cyber attacks are happening.

Parag Agrawal:

I don't know if I'm in any camp among those. I am in the camp that one, I think model capabilities will improve. I think they will improve whether or not you want them. I think. The rate of change is high and will remain high. I think it is likely that the rate of change is going to be disruptive for people. I think it is— those are things that will happen. I don't believe that there should be— I don't think the solution necessarily is rushed regulations. Now, if someone has really thoughtful regulations, I am open-minded. I currently don't know enough to have a hypothesis that great— in this moment of change and new technology with unknown and unknowable ways that it plays. I don't think we can have high confidence in any regulation we write. At least I don't know what that would be, right? So in the absence of seeing, of being persuaded, I believe that allowing people to build, incentivizing people to solve problems, to help, I believe in the Rebel Alliance of all the other entrepreneurs. To go solve things in a way to— if there is a societal problem, I think people will show up to try to solve it, either for profit or not for profit. so, that is my view.

Nakul Mandan:

Cyber risks that are coming up, you believe will be solved through entrepreneurship on the security side?

Parag Agrawal:

Yeah, no, I think they are, right? So for example, the cyber risks are coming up. You're already seeing without a law in place, the people building these models first try in whatever way they can. And you can be a skeptic about their intentions and whether they're pure or not. I like— but they are trying to prepare the world about the kind of risk They are trying to get people to, to find issues and solve them before they become broadly exploitable. They are perhaps compromising their own product experience and use cases to minimize that risk. And they're trying to find some judgment calls on risk-reward on various dimensions, which to me is how things should be. Right? You can't have this rapid change with no potential downside risk, right? But you're already seeing people feel accountable enough to do that. Now you can say they should do more and you can have a reasonable disagreement, right?

Nakul Mandan:

Let me steelman that just for the debate, right? bad technologies have existed forever. All, technology can be used for good and bad. So, guns are regulated, but even driving is regulated, right? Like you have to have a license. You can't be drunk and driving. So why shouldn't there be more regulation here?

Parag Agrawal:

I'm, not saying there shouldn't be. I am saying that a lot of regulation has intended effects and unintended effects. And at least my lived experience is there exists a reasonable class of regulations, especially around technology, that tend to have negative and unintended effects. So now the burden— I'm not saying no to all regulation. I am saying the burden of proof is on the specific regulatory framework we use and how we make sure that it was— it is very likely to have the desired effects and not the undesirable risks. To me, that is the burden of proof. And the proposals that you see out and about, to me, don't meet that burden. That's what I'm saying.

Nakul Mandan:

What do you think about the, social upheaval that might be coming from a jobs perspective? I know it's a speculative question, but where, is your heart and mind right now in 5 years, 10 years? Do you believe Software engineering jobs go away or they just evolve? Is there social upheaval coming or people just need to upgrade their skill sets or change their skill sets?

Parag Agrawal:

I think if this change was happening slowly, I would have taken a— I'm a tech optimist and I'm an optimist in our ability as people in society to evolve. That is my baseline point of view. In this moment, the rate of change and the potential rate of change feels high enough that I don't feel the optimist conviction to sit here and tell you that like, oh, there is like— society will evolve around this technology and everything will be great. Like, I don't feel that way. I think the delta between the AI diffusion is like very uneven right now. Like we kind of here in San Francisco live in like a weird bubble where you're surrounded by people who live in the bubble and kind of believe more in AI. And when you pop out of the bubble, the world is different. Then you go even further, the world is very different. So I don't know, I feel the rate of change is high enough that there is a— I think in the long arc, society will evolve, but along the way, there are likely to be hard periods. And yeah, so I worry about that. I, I don't have meaningful, constructive solutions other than the piece of advice that I would give myself, which is adapt as fast as humanly possible to the world around you. So try to assume that there is going to be this large continued rate of change and try really hard to adapt alongside it and bring the people around you and help others adapt to—

Nakul Mandan:

do you believe the number of software jobs in 5 to 10 years will be dramatically less?

Parag Agrawal:

I don't know how to define a software job in 5 to 10 years. I think the jobs will look different. I also think like I'm in the extreme positive sum game. So I think there will be way more value creation as a result of what you might define. The TAM will be much larger. That I believe. The economy around this will be much larger. Yeah. Now, how many people will do that? I don't really know, right? So in aggregate, the economy around this will be much, larger. Now the question is, will it take way more people to support that large economy? Will it take the same number of people? Will it be like regular growth like we've seen over many decades? Will be shrinkage. Like, I honestly don't have a framework to know.

Nakul Mandan:

Yeah. How are you guiding your own kids for the skills that will be relevant for them in years to come? Right. For you, the curiosity, the math skills really panned out well.

Parag Agrawal:

Yeah, but I think those, like, listen, it's like not hard to believe that you'll need hard work in the future. You will need curiosity in the future. You will need to communicate with the world and other people and models. So those baseline things are the same. I think you— at least I feel that as a parent, you have a role to play in those. You have a role to play in creating the environment and their peer group. But I don't think it's up to me to really decide, like, what they pursue or what catches their fancy at what age, right? And I think like I have evolved from being happy-go-lucky to math to physics to computer science to software engineering to leadership to founders. Like, it's like these are arcs, various times, different things. Kids will evolve, right? Like, I think in, middle school you'll have different interests, in high school it could evolve. Time it could evolve. Who knows what college will be by the time they're there. But should they be curious? I want to make it so. Should they enjoy sport and athletics? Yes. Should they enjoy teamwork in some of those pursuits? Yes. Should they work hard? Yes. Should they enjoy music? I hope so. Or art or something else. so these are like not complicated things, and these are essentially all you as a parent get to play around with anyways, right? But I do think one thing I really feel great like is my kids appear to be hardworking. They probably get that from my wife.

Nakul Mandan:

So last question for you, a young person watching this wants your life, wants the life that you've built. What would you tell them that this journey actually requires?

Parag Agrawal:

Hard work, a great peer group, getting lucky with your— with a lot of things, including family, choosing the right spouse, and only working and getting lucky enough to only get to work on things that you really enjoy. That's not a privilege you can take for granted. But I've gotten a bunch of the ingredients right and made a few good decisions which are consequential, that's what it takes.

Nakul Mandan:

Parag, this was amazing. Thank you so much for coming on.

Parag Agrawal:

Thank you.

Nakul Mandan:

Yeah.