Will Agents Surpass Humans? with Vercel CEO Guillermo Rauch | 🟡🔴🔵 MTS Live

Yeah, to give you an idea, that figure was 3% 6 months ago. It's now more than half. So more than 50% of deployments to Vercel are being made by coding agents, which is insane. It seems like the agents are far surpassing humans. When we look at the data for who's browsing the web, over 70% of traffic to vercel.com on the logged out side is agents. So agents are truly taking over both on the production side and the consumption side, which is obviously really good news for anyone that's in the infrastructure business.

Thank you for having me. It's exciting to be here.

Yeah, to give you an idea, that figure was 3% 6 months ago. It's now more than half. So more than 50% of deployments to Vercel are being made by coding agents, which is insane. It seems like the agents are far surpassing humans, and we don't see this only on the code production side. When we look at the data for who's browsing the web, since we started out with Next.js, a framework for building for the web, we have a lot of data on who's visiting the web, who's using the web. Over 70% of traffic to vercel.com on the logged out side is agents. So agents are truly taking over both on the production side and the consumption side, which is obviously really good news for anyone that's in the infrastructure business.

Yeah, so you mentioned I created Next.js. It was about 10 years ago. It was a bright sunny morning and I realized that building for the web was still too difficult. And there were things like Wix, Squarespace, WordPress. You could publish to the web, but if you wanted to create something that's truly production grade, truly great, it was really hard still. Even with the gift of the cloud and all of the things that were happening, open source software, it was too hard. So Next.js was kind of an overnight success. I had been thinking about it for many, many years, and I had been programming front ends for at that point like over 10 years. But I saw that opportunity: can I bring simplicity? Can I emphasize developer experience, not just infrastructure prowess? And now, 10 years later, we're seeing the same exact opportunity. Eve is what we call Next.js for agents. There's a need to create agents. Obviously every enterprise you talk to, every individual, every startup, they want to create agents. And the options are quite insufficient, I would say. You can marry yourself to a model provider, but you guys were just talking about today, a new model dropped, GLM, amazing price performance, open source. And so we wanted to create a framework for building agents that was truly production grade, really easy to use, and gave freedom back to developers and enterprises. That's kind of what we stand for.

Yeah, on the coding agent side, I think we're seeing a massive proliferation of software being built and we can get into the pros and cons. I think it's mostly positive. Anytime a technology gets democratized, it's better for the world, period. I'm anti-gatekeeping. No, it's only real engineering if it comes from Silicon Valley. No, let's democratize software building for everybody in the world. That's what Vercel was created for. So I'm all for it. But I'll tell you, enterprises are petrified and they're still making a lot of mistakes. They're seeing a massive proliferation of shadow IT where you pick up cloud code and then you want to deploy, you want to build. Do they have the infrastructure that goes really well with those coding agents that can absorb all of this new-found deployment power? And so that's kind of what we announced today with our enterprise apps and agents platform, where no matter what coding agent you're using, there's a Vercel plugin and there's a secure platform not exposed to the internet that you can pair with a coding agent. And that dovetails into the other concern that people have: is all of this juice vis-a-vis token spend worth the squeeze vis-a-vis what are we getting in return? Are we getting high-quality products? Are we making customers happy? On one hand, if you're just wasting tokens and you're not shipping software, you're probably doing it wrong. That's why this conference is called Vercel Ship. We want to give you a deployment platform which gives you a baseline: I'm spending tokens and able to ship code. The other consideration is quality. Coding agents are good at adding, adding, adding, adding. Every engineer knows sometimes the best ships, the best patches, the best software releases are all about simplification. They're all about abstraction. They're about deleting code rather than adding code. So there are still shortcomings in the models. They're a really powerful hammer that sees everything as a nail and they follow what you ask and they follow what you say. And they're starting to lightly push back more. I do a lot of what you can call vibe coding or agentic engineering myself. I want to really test the models as they come out. GLM looks great in benchmarks. I also want to find my own personal conviction on how good models are. One of the things I've been noticing about recent models is that they really push back more and they're better architects. I'm super hopeful that we will continue down this path of increasing quality. Everything that you could say a model cannot do, I would add an asterisk there and say, look, I also didn't think models would be operating at a level of principal or super senior engineer today. And lo and behold, they're getting there.

Really, to your point, the running joke is, and I met so many people that do this, everyone turns on YOLO mode in their coding agents. And for good reason. There's also a lot of hype about loop engineering now and models are being trained on longer horizon tasks and they can do a lot by themselves. So people are naturally trusting them more. But you need a guardrail on the other side. For example, with Vercel's secure enterprise deployment platform, the guardrail that we put in is that by default we don't ship to the internet. We protect it with a technology called Vercel Passport. And we say, if you're logged in to your identity provider of choice, like a smaller startup might be Google Workspace, an enterprise might be Okta, we say, okay, you have to be logged in to appreciate your new vibe coding gem that you just put out. And so we created this sort of guardrail around what the coding agent can ship. You also brought up supply chain attacks. Just today Microsoft broke the news of a massive compromise of the JavaScript supply chain. So we're also working on a lot of what we call AI firewalls. When the agent runs compute, we put the agent in a sandbox and we limit what it can do. We say, hey agent, you're free to operate here, but you can only talk to the hostnames and to the subset of the internet that I allow you to talk to. Because the reality is that in order to get really good performance out of agents, you need to give them some power. You need to give them a computer. The metaphor that I use is: in any sort of modern job, the very first thing that happens after you get hired is you're given a computer. And the same thing is happening to agents. When you create an agent with Eve, we give it a sandbox. We give it a computer. But we have to put guardrails around that computer. We have to limit the network. We have to give the developer controls around what data can this agent access. Can this agent access all of the database? A great use case that people are excited about: data analytics. Will my agent be able to access the entirety of my data warehouse? Well, if it goes even slightly off the rails, you could have a massive breach or data compromise or privacy violation in your hands. And so we're basically weighing this consideration: here's the framework, really powerful, really easy to use, but here's the power to put guardrails around this thing.

Yeah. Starting with: it has to be cloud infrastructure. I think we're still in sort of early days of the internet where I was literally hosting my web pages out of my computer in Argentina. I would host my web server Apache, open it, hope that my ISP keeps my IP stable. When I see people using Open Claw and things like that, during the early days wild west, they're hosting it on their own computers. They're giving it full access or maybe they're getting a Mac mini and whatnot. Now we're entering this new phase of realizing: we have to think about security, we have to think about scalability, and very importantly, we have to think about cost efficiency. So one of the really cool attributes of Vercel Sandbox is that it's serverless. When the agent needs to think, it comes online. When the agent sets up a schedule for itself, it can come back to life. If the agent is not doing anything, it can go to sleep. And we even innovated in the pricing model. Vercel Sandbox only charges you for the actual compute. Because what happens a lot with agents is that they can think and think and think for a long time. I was reading some commentary on this new GLM model and someone was saying, wow, this model thinks for 15 minutes before it writes its first line of code. You heard that right. 15 minutes which is pure thinking happening in the GPU. In that case, the sandbox is sort of idling. Think about the economics of cloud infrastructure. This means that Vercel can create massively efficient infrastructure. It can pack a lot of sandboxes. We can have these massive agent swarms. We can enable enterprises to launch huge swaths of agents in parallel. And basically, you're enabling things that are simply not possible with your own computer. My prediction for the rest of 2026 is we're going to see this displacement of agents. We're going to see them graduate from your own computer into the cloud. They're going to parallelize. They're going to reproduce, for lack of a better word. Companies are going to get a lot more bang for their buck. And again, this means more power to the developer. The ability to scale yourself and sort of commandeer a team of agents will become even more of a reality.

Yeah, we have a simple mental model. There's pixels and there's tokens. And you need infrastructure for both. If someone's telling you you just need the model, I think they're oversimplifying. The reality is that you also need the immersive experiences, especially for the outputs. HTML is all the rage. We announced a very simple feature, vercel.com/drop. You can drag and drop an HTML file. And even before we announced it, this is the kind of crazy thing about agents: agents picked up on its existence before we even announced it. The product was taking off before we even told the world it existed. It's because what models are doing, especially cloud models, when you ask it for something where it makes sense to produce a visual artifact, they produce HTML. They don't just tell you, hey, here's five bullet points and a little table. The rich output is a beautiful thing. So long live the web. We help people that are even less savvy than our average developer audience to just drag and drop a file, get a URL, and it's hosted professionally on the Vercel platform. Super bullish on pixels, but I'll tell you, tokens are growing at a pace that makes pixels look like child's play. The best way to think about our AI gateway, we use a metaphor that it's like a CDN for tokens. When the web kind of got started, websites would go down a lot, they would get hacked, they had poor observability of what's happening, and you couldn't understand traffic. So we invented the CDN. And Vercel CDN today runs some of the, I was joking today at my presentation here in the UK, we host the London Northeastern Railway Company, we host Octopus Energy, we host Co-op, one of the great retail grocery store companies. All of those are flowing through the Vercel CDN. But when they use intelligence, when this software starts acting more autonomously, it needs to get tokens from somewhere. And so the Vercel AI Gateway is like a CDN, but for tokens. And today, going back to it feels like we're in the early days of the internet, tokens are sort of having those same failure modes of the early internet. Anthropic, seeing so much growth that they have struggles in scaling the delivery of their inference. So Vercel AI Gateway can do seamless failover to other providers, like AWS Bedrock. We're also seeing a lot of liquidity in models. People are changing models, they're experimenting a lot. Last week, DeepSeek didn't exist. Now it's 30% of Vercel's AI Gateway token traffic, which is kind of nuts. And so, because tokens are the object of interest now, we're seeing the demand for infrastructure that is purpose-built for agents. And we've been calling this agentic infrastructure.

Yeah, so we're defining the category and we're a very public company, so we share everything that we learn as we build and as we design with our customers and we serve our developers. The very first thing that happened that motivated agentic infrastructure is coding agents. Coding agents will start producing way more software. You mentioned it in the intro. Teams that use Claude Code deploy five times more frequently on Vercel than teams who don't. So you really need infra for coding agents. Think of it as the AWS for coding agents. The other one is: people will continue to build websites, but now they want to build agents. That's why we built Eve. And so the other pillar of agentic infrastructure is that the thing that you're deploying is no longer a website. It's an agent. Last but not least, perhaps my favorite thing for my peace of mind, the peace of mind of the world, is that the infrastructure itself is autonomous. Good example: model goes down. Do you get paged? Or does the infrastructure repair itself? We showed some examples today that are quite astonishing. We're seeing that agents are able to fit so much data and telemetry in their heads. The metaphor for head of an agent is the context window. The context window of a frontier model fits three copies of Harry Potter. My context window fits 10 tweets, it's whatever I'm consuming at this given time. The case for Vercel becoming the foundation of agentic infrastructure is that we see all the data in the world. We see every crash. We have all the client-side telemetry. We have your performance data. We have the runtime data because we're actually running the apps and agents. And so our gift back to the world will be a cloud where you don't get paged. A cloud that optimizes itself. A cloud that doesn't give you so many bad news all the time. I think a lot of us in the cloud industry grew accustomed to staring at dashboards to tell you, oh my god, the API just went down, things are slow. In a world of agents, things just get better all of the time. And again, these agents are working behind the scenes for you.

That's the grind. It's real. And to your point about, or Jean's example of free as in puppies, that's the thing. And I want to be a responsible steward of agentic infrastructure. Software has never been easier to build, but it does have a maintenance cost. That puppy grows up and you have to clean up after it, right? Even though I believe and I'm quite confident that agents will do a lot of the cleanup. You're still responsible for that software entity existing in the world. And so yes, I'm all for more production of software but I do think that quality is still top of mind for me. Quality is still the priority, and delivering great customer experiences for everybody, for humans, for agents, pixels and tokens.

Yeah, so to give you context on what I consider to be the reality of task displacement, not job displacement, those 30,000 questions would have required me to scale the data team. I had to hire more analysts. I had to hire more data engineers. I had to keep growing headcount to keep up with the demand for business intelligence. Nowadays, at Vercel, the way that we scale business intelligence is by building an agent that knows everything about our business. If you just joined Vercel and you need to get up to speed on context about the business, growth rates, example use cases, case studies, you have this all-knowing being that can answer your question. And it kind of becomes addicting, right? That's why there's 30,000 questions for a bit over 700 Vercelians. But I always think of it as an insane competitive advantage that I want to give back to the world because I live it and I see it every day. I see junior engineers or junior salespeople that know so much more about the business because of this agent. So we took everything we learned about building those agents and we poured it into Eve. My prediction is every single company in the world will start by building this base layer intelligence. Usually, when you would start a startup, the very first thing you would do is create the website, register the domain name, get an email inbox, incorporate. I think even before those things, you're going to turn to Eve and you're going to lay the first brick of the intelligence of your business. Because this thing has to build a business with you. It's your first hire. It's not even a digital employee. It's a digital all-knowing Jarvis-style intelligence. And again, building this from scratch requires so much bespoke infrastructure that if you were a startup, one, you couldn't build it. Or you had to take shortcuts that compromise security, privacy, availability. And if you're an enterprise, you're terrified because you hear about Open Claw and you think, oh my god, it's going to get all my data into the internet. And so we kind of delivered on this: okay, we lived it. We operate in highly regulated environments. We have customers that straddle healthcare, fintech, highly regulated services. We're operating at the scale of being behind a lot of emergency responding services. And so we needed to build an agent framework that lived up to that sense of responsibility. This is kind of how we build at Vercel: we first have to build conviction, we have to build it, we have to do it ourselves, and then we share everything we learn with the world. Crucially, we open source these things. So Eve is fully open source. Obviously pairs extremely well with Vercel's generative infrastructure. But we're learning and sharing as we go.

Yes.

I believe so, and I'm building that world. Everything we do is around putting the human first. Eve is a framework for humans to write the intelligence in the agents that'll take them to greater heights. The reflection that I'll leave you with is that we're still in the early innings. We're very much in a takeoff moment. I sometimes look at the news. I try to not look at the news too much because there's just so much debate that can be misleading sometimes. Token growth on the Vercel AI Gateway is accelerating, growing at 50% month over month. Spend continues to accelerate. There's no sign of stopping. There's no slowdown. It can be scary because how far will this go? But I do think we're very much in control. We're building the frameworks, guardrails, secure deployment technology to continue to be in charge. We're betting on model diversity, which is very, very important. Vercel's AI Gateway allows you to choose. It gives you the power to choose. It gives you freedom from this AI that could, left to its own devices, take over everything. And we might end up with one AI for all of us. That's not the world I want to live in. I want to live in a world of choice and maximization of human potential, and that's Vercel's mission to build.

See you soon.