The Agentic Web Powered by Vercel

Well, hello everyone. I'm Jean Grosser, the COO of Vercel, and this is Guillermo Rauch, our founder and CEO. Uh, G, between the two of us, I am pretty sure that we have been on every single developer podcast out there.

Every podcast, and I think multiple times.

Yes. Uh, and so I am thrilled to announce to you all today the inaugural Vercel podcast. Thank you. I'm here from Dreamforce.

First episode ever.

So look, everybody in this room has some sort of mandate and you probably don't yet have a clear picture for exactly what that entails. And I think both you and I spend a lot of time talking to leaders out there and we find that that is pretty common. We recently raised a Series F, Series F. And when we did that, you shipped a blog post out towards the AI cloud. Exactly. Sort of where we saw the future headed. And I think what struck me in that is the fact that not even the giants of technology have it figured out either.

Yeah, AI is like the great equalizer. We're all being disrupted. We're all on common ground.

So we're, as you just said, we're talking about something fundamentally disrupting enterprises, how they build software, how you put together teams, how you think about roadmaps. And we all get to do it together. I joined Vercel because I believed that in the fullness of time, every pixel on the internet could be built on Vercel. And then I rapidly came to also realize that hopefully every token will also be processed on our platform. And I think AI is accelerating that vision even further.

So—

Absolutely. I'm excited to dive in with you. So to kick us off, Vercel is celebrating its 10th anniversary in short order. And the— you founded the company based on a lifelong belief about the dynamic web, and we're now seeing that turn into the agentic web. So tell me a little bit about what that means for the enterprise.

Yeah, speaking of disruption, if you think about the web as it's existed so far, you've thought about domain names, .coms, going to websites, getting snappy responses. Hopefully when you go to those websites, you find exactly what you want at the right time. And so we've been huge advocates of a dynamic web because we all want to go to these applications and get exactly what we want out of them. So we call the first chapter of Vercel the frontend cloud. Yeah, it's all about creating the best possible customer experiences. But I think rapidly, you know, the platforms are shifting. You used to ship applications that were very much pixel-centric. Now you're shipping agents that you might not even see, that you might interact through conversation like we just showed yesterday with Slack or with voice. And so this means there's a new category of software and there's an opportunity to create the best platform for this agentic software.

And you and I talk a lot about the shift from pixels to tokens. In June, Vercel launched the AI Cloud. What is your view on how things are going to shift with the internet and in particular the infrastructure you're going to need to power that?

Yeah, the cloud was born with Amazon Web Services. Web was in the name of the origination of the cloud. If I was to rethink the cloud today, I would be thinking about agents as the primary piece of software that you deploy. And what this means fundamentally is that from the developer experience. You have to rethink what it's like to build and to put in that first brick. You have to rethink about the infrastructure. There's probably new kinds of services, many of which we're already creating, but many yet to be discovered that are all coming and emerging from this idea of building agents. And so AI Cloud is what we call this emerging category of what is the perfect platform on top of which you can do this AI transformation.

And what are some of those components like? You know, we talk about you need different frameworks within AI, you need different compute runtime. What are some of the big things you think will change?

Yeah, as an example, the very first thing that you wanted to do with the cloud for a lot of us was shipping a great e-commerce experience. So Amazon.com was sort of like the, like, poster child of cloud infrastructure, right? And so I would go to this website and like, what I want is 100 millisecond response times and a personalized set of products and experiences just for me. I want the website to remember who I am, etc., etc., etc. I feel like the cloud was under a certain time pressure. You go to a site and it needs to respond really fast. With agents, it almost seems like it's the opposite. It's like being flipped upside down. Instead of the agent responding right away, I might just want the agent to go and like build me something for 3 hours. And then come back with me to me with the exact answer and the exact insight. Yeah, I believe that the AI cloud will be a lot more event-driven. Maybe the agent starts working off of a lead that just emerged from my website and it goes off and does a bunch of compute in the background. And so if you give this problem to developers, it starts to become obvious that they need new tools. Yeah, they need new best practices. They need new cloud primitives.

And so within that, I think one of the hotter topics right now is what becomes of the software engineer 2 to 3 years from now. Are we all just managers of agents? You know, is everyone a developer?

Yeah. As someone that grew up with, you know, barely having access to a modern text editor, I've seen all of the evolutions of software engineering. I believe that the closest thing that we've had to AI was something that Vercel has been known for, which is frameworks. We created the world's most popular React framework, Next.js. It sort of gives guardrails to enterprises to be extremely productive. In fact, I'll say the number one difference between traditional clouds and Vercel is that we took a very holistic approach to solving everything. We give you the framework, we give you the dev tools, we give you the services, we give you the CDN, we make everything really fast. It's like the cloud on autopilot. And so Chapter 1 for us is putting infrastructure on autopilot, the self-driving car of the cloud. Chapter 2 for us is the self-driving car of software itself. Yeah, code is now being generated rather than written by human beings. I believe expertise from software engineers will continue to be highly relevant. What I think is changing is that the very competent developers are now becoming basically agent orchestrators. You know, I, I walk around the office and I see some of our young AI engineers and they're like, you know, managing 5 or 6 agents at a time. It's a little bit like you're becoming an engineering manager, even if you are an individual contributor. But what's even more cool to see is that now if I'm walking from the engineering room to HR, I'm also seeing them generate products with AI. I think what's really amazing about AI for software is that it's the ultimate democratizer. Yeah, frameworks democratize us and got us here. But I believe that AI code generation will have an even more profound effect. I think numerically you can think of, of going from like 20 million software developers to hundreds of millions of software developers. Yeah.

And so that obviously will be a tectonic shift. I think the other thing there's a lot of debate around is if all these agents are just doing stuff on our behalf in the background, is there even a UI anymore? Do pixels matter?

Yeah, I believe so. So at the end of the— at the end of every agent is some outcome that is highly valued to that user. So we have customers like Gamma who are automating the production of slides. And so you're going to do a bunch of agentic work, but at the end of it, there is a presentation. Yeah, we have customers that are automating spreadsheets. So you do a bunch of work and at the end of it there is some result. And the best way to present this result turns out to be a spreadsheet style UI. For v0, you know, again, the agent is creating software for you, but you're actually on the driver's seat kind of saying if you like the design or not, you're— you want a high fidelity user interface for the task at hand. And the other thing that's happening, and you see this with ChatGPT, especially the recent news around ChatGPT being able to embed your UIs and your products within their experience. It's obvious that text is not enough. Yeah. If I ask an agent, what's the weather today? Well, I might want the best possible representation for weather and its forecast in a beautiful, cloudy, foggy San Francisco icon. And so I believe UI will continue to be really, really important. My sort of high-level framework is your own dot-com. You have to continue to invest. You want to make it agentic and AI-powered itself, but you also don't want to not play ball with all of these other agents. And so you want to also show up where the conversation is happening.

Yeah, actually, a really good analogy somebody gave me recently was if you in marketing was when social started being popular. So people were like, you know, either do I not need that at all or You know, is it going to overtake all of the traditional marketing?

Same fear response.

And so it's the same sort of thing of like, of course you're dot-com, you want people to come to a place where you own the brand, you own that experience, but you also now have this new channel in having your UI in ChatGPT. And so of course you want to be anywhere relevant.

And I believe that also from a game theory perspective, I think none of us want to get disintermediated.

Yes.

And to the extent that you can continue to invest in that customer loyalty, you have the domain expertise, you have the brand, you have— you understand your customer the best.

Yeah.

And so if you can— and this is the promise of the cloud— build AI capabilities into your products and into your front ends in addition to integrating with the agents that are becoming overwhelmingly popular.

Yeah. So you talked a little bit about v0. We demoed v0 in Slack 2 days ago at that keynote. And so it's really the view is there you're letting everyone create at the speed of ideas. So do you think this, you know, what is this going to enable for different customers out there that maybe now non-developers actually can contribute to software development and your website?

Yeah, anytime I've shown v0 to people, if I show v0 to someone who's never seen it before, they have such a shocking response. It's like, well, this is the end of all software because I show them how you can prompt and generate the application of your dreams. But one of the first responses I started getting, especially from non-full-time developers, solutions engineers, started thinking about, well, when I'm in the process of selling a product, what can I show the prospect? I can show them slides. I can talk, I can yap, or I can show them that I'm already doing the work. I can show them an artifact that is, you know, demonstrating the capabilities of my technology. So this is what we did when we helped Overstock migrate over to Vercel. Our solutions engineering team took the existing Overstock, which was hosted on a legacy platform. I'm not going to name names. Legacy platform. And they took a screenshot and put it into v0, and v0 started porting the platform into cutting-edge Next.js and Vercel code within one phone call, right? Within a Zoom call, you could start to generate this migration. And so v0 as a catalyst, not only for new ideas, but for transformation has been an incredible use case that led to a conversation with the CEO and president— or sorry, president of Overstock and Bed Bath Beyond, Marcus Lemonis, who's somewhere here in the conference. And he started thinking about even generating software himself. Yeah. So he showed me how he was using— at the time, I believe models change every day. I believe ChatGPT had just come out with image generation. And he asked me, well, I can use ChatGPT to generate an image. Could I use v0 to generate software? And we started going back and forth on creating a landing page for Camping World, another one of his businesses. This man is super prolific. And so v0 is starting to change how we communicate, and it's changing the conversation from just being words to being prototypes of software and even finished production-grade applications.

You have to— along those lines, you've got to tell the story of with Camping World and LA fires, because there you have a situation, LA fires are happening, rapid response. It's so hard for that type of company that doesn't have tons of software developers like you have to be able to respond and spin up a new site.

And it's fascinating because now you have here someone in leadership that wants to make a difference for the world. He wanted to voice out Camping World can help with this disaster relief efforts. And again, we're able to generate from English prompts. We were able to use his brand language. We're able to use his iconography. We're able to use his colors. In fact, something that was fascinating is the model already knew about Camping World. So all we had to prompt NVIDIA was make it look like Camping World. And so it was, it was just an amazing experience.

I mean, I think you sort of hit on this too. One of the things I've loved in talking to different executives that I engage with is they basically get to go back to— I used to have to try to like articulate an idea, hope my team would then go write a PRD, put together a prototype, etc. I was talking to the CPO of Avalara, so tax software company, and he went from prototype to patent. So literally, yeah, had this idea of like a thing that he wanted to do for a different way to integrate a bunch of different esoteric tax systems. And he literally, the night before a board meeting, gets on v0, natural language, just explains what he's trying to do. It's a very complicated workflow and takes this prototype, shows it at the board meeting the next day, and then within a couple of weeks they'd actually built it full production ready and then have since gone on to patent it.

I mean, no pun intended, this really is the power of computers. I've been hearing a lot of feedback, customers like Block that they struggle with doing whiteboarding or prototyping of numerical problems. Yeah, because you really need UIs that have real data and can demonstrate interaction and input and do calculations and adapt. We talked about dynamic web. Yeah. You know, what is ahead of dynamic web is what I call the generative web. So if you can use v0, you're not just designing a static sort of, you know, blueprint, you're actually bringing software to life. And yeah, it's accelerating everybody.

So the other big thing you talk about constantly at the company is, you know, like the last decade plus has been developers dealing with a lot of problems. You are constantly getting alerts, you have tons of dashboards, data, and that with agents you can actually— yeah, you get paged in the middle of the night. And so with agents, we can now actually go to sleep Solutions, not problems.

Yeah.

So what does that look like for you?

So you worked at Stripe for many years. Yes. 10 years, which is like 1,000 years in software engineering. So thank you. Patrick has this really funny meme of anytime he hears ducks, like duck sound, he gets PTSD because his pager sound in the early days of Stripe—

The duck quack.

Was duck quacking. Like quack, quack, quack. It's like, oh no, that means that the Stripe servers are down.

Yeah.

So over the past 10 years, Vercel has been working relentlessly to make outages and managing servers and infrastructure a thing of the past. Yeah, your website just scales, it doesn't go down, it doesn't matter if it's dynamic, static, etc. But in reality, you know, the, the world is, is changing very, very rapidly. So the, the, the most— I think most people are still stuck in, in data centers on-prem, like to say something like this sounds like a dream. Oh my God, I don't have to worry about infrastructure anymore. In reality, I think a lot of us have felt a little bit hostage to this idea of the cloud is something that just gives me problems and problems and problems and problems. I need to set up dashboards, I need to set up pagers, I need to set up alerting, I need to set up alerting on top of the alerting on top of the alerting, etc. And I believe that the ultimate promise of agents applied to the problem of the cloud is that they just give me solutions. They just give me good news. And so the best example that I have is that when you roll out bad code, we're all used to the fact that some metric might start trending in the wrong direction, right? Errors go up. The typical solution is we page a developer.

Yeah.

If we have everything perfectly set up realistically, maybe nothing happens and start losing money, which is devastating. And so the idea that the agent can be the one that's on call is what I mean by the other meaning of the AI cloud is one where agents are always self-healing your software, self-optimizing your software, meaning even performance optimizations can be done by the agent sends me a pull request. And ultimately even suggesting new actions to be taken. The classic example here is Vercel deals with a huge number of threats and denial of service attacks every day. Some of them are highly targeted on individual properties. And so what we can do is we can have the agent perform an investigation on the anomaly of traffic. We actually enabled it. We announced this yesterday. Where an anomaly gets detected and the first response is an agent that investigates the available data. This is the ultimate promise of the agent cloud.

Yeah, I love that we, you know, put— we have some pretty solid security engineers that typically investigate this and you put them side by side with the agent and it's magic. Agent wins.

The typical response from the security industry has been you pay professional services and professional services and professional services and you hope that they can solve your problem. Yeah. Is there someone laughing in the audience?

Agree to agree over here. They've been there.

I believe that we can solve this with tokens. I believe that we can solve this with agents.

Yeah. All right. So segueing into what other agents do we think might be useful out there? You know, we're, we're constantly talking to companies that are trying to figure out what to go, go ahead and build, maybe cover a couple of things we've been building internally that we've seen the most ROI from?

The obvious thing that everyone will say, and I have a nuanced perspective on and I have firsthand experience on, is support. Yeah. So number one, one of the product development philosophies of Vercel is we're always patient zero of our technology. So when the web was becoming the agentic web and ChatGPT came out, we started building our own agents. One of the first agents that we built was our customer support agent. And I have to say, I was a little— I had some trepidation. Why? Because our customer base is highly technical. Yeah, they want— they come to us because they want the experts of the web to be on call. And so I wanted to make sure that this support agent was going to be extremely high quality. It's not just a cost optimization. We give better answers faster. And so fast forward to today, our support agent answers 75% of customer tickets. And these are highly technical questions, right? Not only that, but CSAT, NPS, all of the signs, vibes, people that tweet at me on X, everything points to the fact that people are happy with this solution. So qualitatively and quantitatively, support agent has been overwhelmingly successful. Obviously, we built v0, one of the largest coding agents on the planet. We built Vercel Agent. I'll say one more thing on support. I had a great conversation recently with a CIO that's kicking the tires on the Vercel AI Cloud, and he was saying that beyond support, what he wants to solve with Vercel is the root cause of why are people filing tickets and calling the call center to begin with. This is— I think it was a little bit of an epiphany in the sense of like we're all jumping to buy and sell support agents. Yeah, but how about customers not calling to begin with? And so we're doing analysis of call transcripts to understand root causes and distribute them to the corresponding teams. Yeah, I think you have other examples on your side of the house.

Yeah, I mean, one of the things I love joining Vercel, 7 months in, um, was I think like week 6 I'd go to Guillermo and I'm like, so I'd like a GTM engineering team. Obviously he said yes. Done.

Um, he was all my engineers.

Yeah, exactly. So we have this like, you know, crack tiger team of GTM engineers, and the first thing I took a stab at was, um, our inbound, uh, contact sales flow. Um, and it was a similar thing, like everybody knows who's worked in sales, you want to be instantaneously responsive, like the drop-off if you if you don't schedule that meeting within minutes of getting the, the exponential decay. Exactly. Um, and so we took one engineer who wasn't even dedicated, probably worked on it 25% of his time. He shadowed our highest performer SDR, looked at everything, um, he was analyzing, the different tabs you go to, data he looked at. And then we turned that into an agent, um, and it's fall— it's got workflow, it's not entirely deterministic. And we had 10 SDRs previously handling this volume. We now have one and the conversion rate is equivalent, but we get to that conversion with far fewer touches.

I want to emphasize how smart— when I heard you did this, I thought it was so smart because the way that thinking models are being trained today, like we went from models that were instant, like the first GPT, to the models that think. The training data for the thinking is the world's experts in each question and how they would attack the problem. Yeah. So by shadowing your best performer, you essentially created frontier training data for a novel problem that is not going to exist in the model because the model doesn't know how our best seller sells. So you basically trained your own model to solve this problem. I thought it was fantastic. And now I'm taking this formula and applying it to everything. If I want to performance optimization, I know a few things about optimizing web performance. You know, our job is to basically, you know, take ourselves out of the job is write down here's the 10 things that I would do to optimize overstock.com.

Yeah.

And now we put that into Vercel Agent and now customers can buy it.

Yeah. I've got to say, like, to put a cherry on top of the lead agent story. So we have 10 10 people, you know how much SDR gets paid. So it's now 1. The other 9 now work on outbound, so moved up the value chain. And it costs us about $1,000 a year to do this on Vercel's AI Cloud, which might not be awesome for me in sales, right? But for anybody else who wants to give it a try to build one yourself, highly doable. So I mean, what is your view too of like how, how much are Should people be getting off-the-shelf agents versus thinking about domain-specific proprietary agents?

Yeah, it's so interesting to share your story because you're COO, you're not running engineering teams, and you saw— because you live in Silicon Valley, you saw every pitch and every billboard of an off-the-shelf AI SDR.

Yes.

And none of them worked. Yeah. And so, and you approach the problem by saying, I'm going to use Vercel, I'm going to build it myself, I'm going to take 25% of someone's time. And that's really all it took to create your custom agent. Yeah. And so I think for us, my mental model is— and this is provocative— but there is AaaS, which is agents as a service. You heard it here first. Agents as a service. I believe that's going to be very powerful in the future. You can just buy them. This is the example of ChatGPT Enterprise, right? You buy them, you give it to your people, immediate returns. But the reality is that they can only go so far, and you should not be afraid of building your own. Use an AI cloud like Vercel, build it yourself, and this is going to be the custom agent world. And I think the custom agent world is now split into two factions. One is— and this reminds me of the beginning of the web, and it'll date me— you can use visual builders for your agents, which reminded me of Microsoft FrontPage '98, which is the first web builder that I used. The promise was, oh, you're just gonna do it visually. It's gonna be great. Well, guess what? We're not using FrontPage today to build high-quality software, right? And so I'm not gonna diss on visual building for agents, although maybe I just did, but I think you're gonna have to, if your business is gonna run on agents, you're gonna have to treat them as, a rigorous software engineering project with CI/CD, Git, auditability, governance, observability, so on and so forth. And so that's what we're giving you in a box with our Vercel frameworks and AI cloud infrastructure.

So one of the things we obviously spend a lot of time talking about is like wave one of AI was assistants. We're now really in this agentic world, and then you'll ultimately be in the autonomous agent phase. And one of the things I love about working with you is you're just constantly out on the front lines talking to customers. Maybe tell me about a customer where you sort of— their workflow and the agentic experience you were trying— they were trying to create were just like, infra as it exists today cannot do this.

The most astonishing example that I, that I've talked to is a company doing IV treatment approvals in hundreds of clinics in throughout the US where the process is fundamentally so complex, most of which involved paperwork. And when I saw the complexity of this agent, I was bewildered because I fundamentally learned that you could not even solve this problem with traditional pixel-based solutions. Meaning if they had tried to help this patient with a better UI, it, it would've just been yet another tool.

Yeah.

Okay, let's train every clinic in the US with this new tool. It's going to solve every problem. And this is where we saw agents just work amazingly well and they are piping every modality of agents. So there is an aspect of voice because they need to make outbound calls. There is computer vision for document processing. There's a lot of intelligence on demand like the traditional ChatGPT style. And the biggest learning from us on the infrastructure side was that in order to support them, We needed to have infrastructure that allows long-running compute that is composed of multiple discrete steps. And the goal is to make this highly, highly reliable and auditable. This is a class of software that most engineers are not used to building. Again, we come from this rapid request-response world of, again, Amazon.com. And so now we have an opportunity to create this workflow-based infrastructure and the frameworks that make it really easy for— I mean, this was built by a startup that has basically gone from $0 to $20 million in ARR in record time. The demand from the consumer side is, you know, very well-proven. 800 million active ChatGPT users. My aunt was texting me about chat today. So the appetite for simplification on the consumer side is enormous. And so now there is a supply problem. There's not enough, you know, infrastructure and developers that can fulfill the level of quality that these applications require. And so this is where we come in.

So everybody comes to Dreamforce because they want to come up with ideas that they actually can go home and apply. So what would be your suggested sort of next step for folks listening?

There's very easy things that you can do to bring agents to life. So yesterday with Slack, we talked about v0 inside of Slack. So if you want to experience the joy and wonder of generating software with friends on Slack, install the v0 integration. It's the easiest path. But that'll also open your eyes, hearts, and minds to how easy it is to bring an agent into where people are already doing work. So Vercel put forward a template for creating an agent on the Vercel AI Cloud that integrates with Slack. So it's very easy to code up an agent, deploy to Slack. And last but not least, we also published a template on how you can integrate with ChatGPT. So ChatGPT recently announced you can bring your own product UI into ChatGPT through this nerdy protocol called MCP. We put a template together. In a couple steps, You just bring your UI, your product, your data again into where the conversation is happening. So those are 3 easily actionable things that everybody can do tonight.

Awesome. Well, congrats. We have just wrapped up our first, you know, Vercel podcast.

Tune in for the next one.

Yes. And hopefully everybody got a couple good insights into the future of the agentic web.

Thank you, Jean.

All right.