0:00Thank you for coming.
0:00Throwback, by the way.
0:01I've been in this position before, in this stage before, and it's a wonderful venue, it's a wonderful time to be building.
0:07So, thanks for having me.
0:09You know, we're both outsiders, right?
0:11You're from Argentina, I'm from Brazil.
0:13And I feel like sometimes as outsiders, we feel like other people might have more advantages than us when we're building.
0:20Can you tell me a little bit more about your journey as you're building and thinking, like, yeah, like I can build a generational company?
0:28First of all, I'm constraining myself not to make a joke about Brazil versus Argentina.
0:37Look, I think what's wonderful about San Francisco is that we're either all outsiders or none of us are outsiders and we're all insiders, right?
0:48There's the two perspectives.
0:49One, that, I mean, you basically drive from the airport to the city and you immediately feel at home if you're a builder.
0:58Maybe too much at home because you see 20 different AI billboards.
1:02And you can kind of put together your entire stack.
1:05Oh, there's Vercel over there, and there's a Stripe over there.
1:08Like, by the time you drove from the airport to the city, you have a startup, right?
1:13Especially with agents.
1:14And there's the argument that we're all outsiders in the sense of someone has said that in SF, the social status currency is how well your startup is doing.
1:27I feel like every day there's this competition and drive and intensity that comes with SF.
1:36And I feel like we're all equal in that sense, that we all have to build great shit.
1:42And yeah, I think it's been amazing.
1:45And I highly recommend people come here.
1:47I think, you know, the other thing that's amazing about displays is the access to capital, both human capital and dollar bill capital.
1:57And I think, you know, having had the opportunity to be here and basically be a nobody other than my open source background and have access to those things was revolutionary.
2:07And when people think about how to reproduce this phenomenon elsewhere, I always remind them that you have to have those things in place.
2:15It's not just the nice weather and AI billboards.
2:20I want to talk about Mood Tools too.
2:22How did that shape the way you think about open source?
2:25Like, open source is so different nowadays.
2:28I was just kind of over here.
2:31I was on X Slack and listening to the previous talk.
2:34And there were a few really insightful things that made me look up.
2:38I think our human attention being such a critical, precious resource is a big part of that.
2:44I also think this idea of taste, the idea of taste has sort of been overplayed a bit recently, but there's something that when I look back on what Mood Tools changed my life, and again, for context, Moodools was this JavaScript library, right?
3:02But more importantly, it was a small community of extremely dedicated people, passionate about craft and engineering.
3:11In fact, the founder of Mood Tools, by the way, we didn't win the popularity contest because the library became more popular with jQuery.
3:21I actually cave that are using jQuery myself too.
3:25But what Moodools really had that I still massively value to this day is Valerio, the founder, would sweat every pixel.
3:36Something I think we share with recent Vercel and a few other companies here in the Valley is that everything we put out is like an art project.
3:44It's not just, you were talking about our partnership with Mercedes and what binds us with companies like Mercedes is we care about the engineering and we care about the beautiful presentation, whether it's the leather interior of a car.
3:59And if you think about the digital metaphor of that, if you were to say today, I want to build the next Mercedes, well, you would have to pay attention to those, to every detail.
4:09You know, it's interesting that when you order some of their cars, the engineer that worked on the engine will sign the engine.
4:16There's a human connection to the engineering and to the design.
4:21And I think you have to create a digital equivalent of that.
4:25And Moodools was the first project that taught me that.
4:28Everything was carefully crafted and curated.
4:32And how do you scale that sense of quality and craft?
4:36Because how many employees at Vercel now?
4:41That's really hard, right?
4:42Like keeping that bar high.
4:44Can you tell me like what are the mechanisms?
4:46Like how do you make sure that that's still there?
4:49Yeah, so as you guys know, and as the introduction said, I've been a framework or tool builder my whole life.
4:58In fact, when I was a kid, people would hire me to work as a freelancer on their projects.
5:04And in order to complete the task, I would first build the tool to complete the task.
5:10I was like skillifying hardness engineering agent, Lube Engineering, before AI, in the sense that, like, I was taking a step back and say, hmm, how could I automate my life?
5:20How could I create a dev tool that is more pleasant to use almost as the side quest?
5:27But I was always trying to like marry the side quest and the main quest into my delivery of the work.
5:32And it worked really well.
5:33I was really proud of my freelancer ratings.
5:35Like, five stars, like, I had hundreds of five stars or whatever.
5:40And so, when you think about frameworks and the reason that I created Next.js and AICK and things like that, they give people a blueprint and they give you guardrails and they give you shape and they give you...
5:53Now I think what's really exciting, and we have Andrew here, skills plus linters or skills plus compilers, skills plus verifiers are kind of the new frameworks.
6:04But what all of these technologies have in common is that they give you, again, that sort of blueprint so that you don't start with nothing.
6:12And when I started Vercel, I realized that I was building both the product and the technology, and I was also building that blueprint for what it meant to join the company and how you should behave within the company and what the rituals are within the company.
6:28So anytime you're building a company, you're building the product that you sell your customers, and also you're iterating on the engine that is your own company.
6:37And so you're kind of working on, I was working on two frameworks, Vercel, the innovation company framework, and Next.js at that time.
6:45Obviously, this was not that clear to me at that.
6:49I thought I was just working on infrastructure and developer tools and trying to ship as fast as possible, as good as possible, right?
6:55But then that was the big insight.
6:57The way that you scale quality is still through technology.
7:02And by the way, technology using AI speak is load-bearing there in the sense that it's technology in the sense of automation.
7:11But there are very simple technologies that are really effective.
7:14Like you onboard an engineer, a designer, and you teach them your values.
7:18And it's a little bit like the principles of the framework.
7:21Like value at Vercell is dig deep.
7:24It's a value that conveys, we know we're gonna have, we're gonna sweat every detail.
7:28Like sometimes I see our designers like nerding out on today.
7:32I saw Wits from our dashboard team explaining some of the magic of dragging a sidebar, dragging the Vercelle dashboard sidebar and the amount of thought, care, and love that went into that interaction.
7:45Which, by the way, you know, people ask me all the time, does this matter?
7:50Agents are using the internet more than humans.
7:51Machine-to-machine traffic on Vercell is 60%.
7:55The human-based interactions matter tremendously still.
8:00Actually, you know what's interesting is I think we all just need to do more work now.
8:05Good or bad news, I guess.
8:07We have even more jobs to be done.
8:09We have to create great agentic interfaces.
8:12Recent, I think, has been doing a great job there.
8:15Everything like copy prompt, markdown, AO, all those things.
8:18And we still have to create great human interfaces.
8:21And so the design team at Vercell kind of embodies this and is almost like our presentation card or a cover letter to the world.
8:33If you were to put this in semantic versioning terms.
8:39What is the version of Vercell today?
8:43Well, I hate breaking changes.
8:46So, I guess we're at 0.0 in a huge number.
8:52Yeah, and I mean that, by the way, like when you work on infrastructure, one of our operating principles is never break broad, never break customer trust.
9:01You know, we recently onboarded our head of product for compute who used to work at Google.
9:08And one thing I loved about getting to know him was he shared that passion on like something that I deployed 11 years ago must still work in some French English accent.
9:19And I thought, it's like, wow, I found my better half.
9:22Like, that's so cool.
9:23Like, I think the same way.
9:26And so there's an element of incrementality.
9:31But there's also, you know, radical refounding or transformation moments that you have to sort of layer on.
9:40The way that I think about it is that you have a foundation, and then you're always working on that foundation.
9:47I think we talked about this in the early days of Recent, where, look, communicating to customers, like that foundation of the company, it's going to define so many different innovations and channels.
10:00I don't know that today would be sitting, we could have predicted back then, oh, yeah, agents are going to need mailboxes and new protocols and MCPs and whatever.
10:10But you do know there's a foundational system that enables packets to be delivered from point A to point B and envelopes and emails and messages and channels.
10:21So we're always working on those foundations, but of course in the kind of like infrastructure sort of hosting equivalent in our world.
10:29But then you layer on this radical new chapters.
10:33So I'd say chapter number one was front end.
10:36Chapter number two was probably AI and sort of AI SDK and a lot of these innovations.
10:42And chapter number three, I do believe is like back ends and agents.
10:45Meaning AI was really interesting because everything has happened so fast ever since ChatGPT came out.
10:51But AI in the beginning was still a little bit like a request response.
10:55Like people would layer it on existing applications.
10:58Like you would open Gmail one day and you're like, they throw AI on top.
11:02They're like, you can press this button and it will refactor your message.
11:07But the third shift towards agents is way more profound.
11:12Because we're letting the compute run for a long time, almost like unsupervised.
11:19And the paradigm that you use to build that kind of software is completely different.
11:25It's more eval-centric.
11:28It's even more non-deterministic.
11:34The human is there to sort of set up the guardrails, but you're sort of leaning into the autonomy of the system.
11:40And so it's given rise to a whole new set of services that didn't exist before, like the gateway, and we just launched Vercel Demons.
11:49We had to create a lot of new, really cool.
11:52And when this happened in the industry, these platform shifts happen, it helps you innovate.
11:57So I would say to your question, like maybe this third major version of Vercel is quite profound because it helps us define a whole new category.
12:06We call it agentic infrastructure, which is the infra that you use to ship agents, but it's infra that is in itself autonomous, which is a little scary sometimes because it's like, oh wow, like I'm letting the cloud run itself, but obviously it's in the service of providing the best possible quality of life and service to people.
12:26I'm curious about how you see that framework layered these days too.
12:29Next.js turns 10 this month.
12:35How do you see the framework now versus before?
12:39I think in many ways frameworks will always be with us.
12:44I think they're transforming and evolving.
12:47Agents still need to use things.
12:51The metaphor that I like to use is, let's say you ask the agent to center a div, as one does.
12:57It'd be weird if the agent says, well, in order to do this, I first need to create a new computer from scratch.
13:03I'm going to order it from Gen Zen, and then I'm going to work on a new operating system and a new kernel and a new web browser, a new rendering engine.
13:11And then I spent 2 trillion tokens.
13:13Here's a red rectangle in the center of the screen.
13:15That would be weird, right?
13:16And so I think there is common tooling and infrastructure that we're going to use as an ecosystem to continue to collaborate.
13:24I think value shifts, though.
13:26For example, protocols today are way more exciting than in many ways that some of the libraries that you would incorporate into your projects because even libraries can sometimes be a liability.
13:38Like I ship a new library and you probably saw there's zero days every single day of this week.
13:45So protocols like MCP, A2A, UCP, MPP, all of these things that allow us to collaborate or make systems cooperate at large scale become really, really important.
13:58The recipe for creating software and combining it with building blocks and verifiers feels like the new framework to me.
14:07So, just to state that again, right?
14:09We do need a mechanism for sharing recipes.
14:13So, I can say, look, if you're trying to build this thing, here's a skill.
14:18Now, for my previous statement, it'd be really weird if that skill tells the agent to, hey, rebuild the entire universe.
14:25It's likely that you're going to be reusing building blocks.
14:30Building blocks, I think, are still distinct from frameworks.
14:32Frameworks were like thick, heavy, very opinionated.
14:36Building blocks allowed the agent to almost like create a bespoke framework if needed.
14:44There's a mental model that I've used a lot, which is scaling up and down.
14:48You know, if you need to create a simple website that says hello world, you probably just want to ship an HTML file that says heading one, hello world.
15:00If that program evolves and becomes very sophisticated, you're going to want to have some kind of framework and building blocks that allow you to express that sophistication.
15:10So that means to scale up.
15:13The ideal technology, and this is why I think coding agents are so exciting, is one that scales down and hits that global or local maximum.
15:22And as your project evolves, allows you to hit that other maximum, meaning that there's no trade-offs whatsoever.
15:29You can be really sophisticated, and that means more code and more load and more building blocks.
15:35Or you can be very raw and vanilla.
15:38And we really haven't had a technology that allows us to express this in the past.
15:42You would always be, frameworks would present to people, a really bad trade-off.
15:48You either had a framework that was too simplistic and said, yeah, it's three kilobytes, gzipped, use mine, it's super cool.
15:58React is 70 kilobytes.
16:00And then you try to do anything interesting, and the whole thing breaks down.
16:04On the other hand, you have Java 4.0, Enterprise, Java Beans, Oracle, JDK, SDK, whatever, and it can do everything under the sun.
16:14And then booting up an application takes two hours.
16:17And so I think developers have always been found in this weird, cornered into these weird choices.
16:23And I think agents plus building blocks plus the skills and verification loops are going to basically rid us of any kind of trade-off.
16:32And I think I also predict what I mean by verification loops is the rise of the compiler.
16:38I've always been a compiler guy in the sense that I've always loved a good build process that gives me a higher level of confidence before the software goes out.
16:47I've also always been an end-to-end test guy.
16:51You know, one of the operating principles or things that I would like to teach engineers at Vercel is like, don't fall in love with the code, don't fall in love with the type system, don't fall in love with anything but a great working product.
17:03And end-to-end tests, for me, were always that technology that allowed me to say, like, oh, cool, you're proud of your, you know, IDRIS-dependent types?
17:11Let's have a thing, click around, smash it, restart it five times, turn off the internet connection, turn it back on, and see how your program behaves, right?
17:21And so that is going to be a very important part of the new agentic software development lifecycle.
17:27And something we're investing quite heavily at Vercel because it's going to allow us to solve the final problem of computer science, which is that you have one million unreviewed PRs sitting there.
17:38It's crazy how the number of draft PRs or open PRs increases, I'm sure, for everybody here.
17:44Yes, and so you're going to need a lot of verification technology more so than framework technology.
17:50And so, again, you need these tools, it's just that the attention will shift.
17:55I feel like it's fascinating as a builder to learn more about how people inside companies are creating software that solves an internal problem.
18:06Can you tell us a little bit about what are you guys building internally, maybe replacing existing subscriptions?
18:13And what are you saying no to that you're like, no, this is something we should not rebuild?
18:19I recently wrote a message to the company saying, here's my framework for build versus buy.
18:27And then with the help of one of my executive team members, we extended that framework to say it's not build versus buy, it's also build, buy, simplify.
18:38So in the past, you could buy software and you could hit this local maximum of like, I buy five tools that solve one.
18:49They also kind of aspects of the same problem.
18:51This is a good opportunity to shut them all down and simplify.
18:56And you can build something that really solves the business problem.
19:03The other one is there's definitely a temptation in shiny object syndrome right now to building for the sake of building.
19:14And there is aspects of a technology that you cannot rebuild easily.
19:19For example, trust, reliability, security, like these are not things.
19:25Long-term maintenance, IP reputation challenges, deliverability challenges, like all of this stuff, I don't want to deal with that problem.
19:35So I try to send a message of balance, not of like, just rebuild everything under the sun, or the opposite, like, oh, don't waste tokens or data center water, whatever, like, just reuse everything, right?
19:50Like, there's a balance in everything.
19:52And so I try to hit that balance.
19:54That said, so I'll tell you some of the cool things that are being built.
19:57I was looking at our design channel this morning, and one of our engineers built basically a design tool that allowed him to do this stress testing of UIs.
20:11So he knows that I do it.
20:14And he's like, okay, could I codify G into a system?
20:18And I was bewildered.
20:20I saw this, imagine a robot that just smashes all of your permutations of your UI.
20:26So this is sort of like the quote-unquote factory mindset of like the job is no longer to like do the pixels on the screen.
20:34The job is to work on the machinery that produces the pixels on the screen or the or the machinery that produces the PRs.
20:41So one layer of abstraction higher.
20:43And I've seen a lot of innovation in this space.
20:47And now what we're doing is the things that work really well that we've discovered, we're going to productize.
20:53So just like we shipped the deployment preview as a thing that customers really, really enjoyed over the years.
21:02Now, we're going to bring to the market: like, what are the things that like imagine, like, what is the next deployment?
21:07What is the thing that allows me to put my software delivery pipeline on autopilot?
21:13And a lot of that was born out of that internal experimentation.
21:18So, I encourage companies to experiment because from those experiments, you'll get your next products.
21:24And I hear this from a lot of customers sometimes that tell me, you know, they get Vercel, they start developing internal tools, internal applications, and then one of them hits hard, and you can productize that.
21:36It's actually, it's like the secret sauce of Vercel is that we experiment internally.
21:42It's like breath-first search.
21:47Like, there's lots of flowers bloom, and then some of them need to be shut down, which is hard to do sometimes.
21:55It's like everyone loves their little pet.
21:58But you need to, and then you productize.
22:01Because otherwise, if you have a closed-loop system where all you do is build internal tools, well, like, who's giving you money for the internal tools?
22:09And so, like, you have to find that balance as well.
22:11Like, revenue, like, top line and then bottom line, and like, it has to save on costs, which this internal tooling can do.
22:18But also, there's the potential of selling it to customers, which is exciting.
22:23Well, I have a rapid-fire session that I want to go over with you.
22:28I was freeze with it.
22:31First question is: what's the last thing you vibe coded?
22:36I have a really good one.
22:38So, we're going to unveil a chess sponsorship.
22:43Our first one ever, chess player.
22:45And I ordered a Bluetooth chessboard.
22:50And I vibe coded a CLI to control it.
22:56And so he's going to do this demonstration where he's going to play against AIs.
23:01We're going to see which AI is best at chess.
23:04They're all kind of bad for what it's worth.
23:09And it was amazing to do this because people are walking around the office and seeing this chess.
23:15It's like Harry Potter.
23:16It's like this chessboard moving on its own.
23:18Because the agent is cooking and trying things in the background.
23:21I'm not even paying attention to it.
23:23I hope the chessboard is still there when I come back.
23:29With DeepSeek, by the way, not to shield awares, but we just announced Token Pass.
23:35And I've been amazed.
23:36So I use this to like, I try to test our technologies as much as I can.
23:40And I was bewildered by the latest DeepSeek model.
23:46Like, it can write any code.
23:49And it's extremely efficient.
23:51Like, it costs cents to do anything.
23:54It's very impressive.
23:56What's your favorite model right now?
23:59I'm, you know, every week I'm sort of like resetting my priors.
24:05I'm really excited about decision models.
24:08So many use, I'm sure you guys as well.
24:10So we introduced the Decide API on AISDK and AI Gateway.
24:16I think we're about to announce reviewing change logs on the way here, and we're about to announce another decision model.
24:21There's like now there's like dozens of decision models.
24:26You've probably dealt with the problem of like when you scale up a platform, you get a lot of like attention, sign-ups, sales, whatever, like all a lot of volume.
24:37And so how do you cut down, prioritize, filter things at large volume, large scale?
24:43Like LLMs were not good for that, too expensive.
24:46So really excited about the Jev style category models.
24:52Yeah, I think, and then I think Opus 5.5 crushed, I think, is incredible.
24:58So combining Opus 5.5 as the planner with sub-agents is something I've been experimenting a lot with.
25:04I'm trying to recreate the TypeScript type checker in as few tokens as possible, or in as few dollars as possible, which is an experiment.
25:13And I'm excited about kind of the, someone was speaking about this, like agent swarms is going to be more interesting.
25:20So think of it this way: like, have you heard of the EDM of like, if you want to go fast, go alone.
25:27If you want to go far, go with the team.
25:29I think the same applies to agents.
25:32You could have the fabled 5.2 million doing one task, one thing, or you could have swarms of agents.
25:40And so I think swarms of agents are probably the next big unlock, rather than people saying, oh, I heard about this new model.
25:47It does feel like we're hitting to like marginal improvements on like Opus 5.5 versus 5.6, whatever.
25:54Whereas I think fleets of agents can solve millennial price mathematics problems or find zero days in PHP or what have you.
26:03And last, and this is the most important one: Messi or Ronaldinho.
26:14I saw him recently at USOB.
26:16I always like that guy a lot.
26:18But jokes aside, Maradona.
26:22Well, thank you so much, Guerrero.
26:23Thanks for having me.
