Moving Fast Without Breaking Trust | Dreamforce 2026

Welcome everyone. Welcome Guillermo. If you know Guillermo, he also goes by G, but I'm gonna practice my Guillermos this whole time, if that's okay with you. Perfect.

Thanks for having me.

All right, move fast without breaking trust. Not easy to do, but let's start with a story, one of your favorite movies is Pixar's film Ratatouille.

Yep.

Yeah.

Still to this day.

Well, let's start by like, why do you like that movie?

Oh my God. There's so many dimensions. First of all, I just love, especially in the age of AI, I heard in the previous fireside we talked about quality.

Yeah.

I think something that distinguishes some of the finest creations, some of the finest pieces of art, movies, everything is just quality. I think Pixar really embodies that. It's something we strongly believe at Vercel. We're just chatting a lot about how, with Paul from Salesforce Ventures, we're talking about the killer app of AI, in my view, is coding, is code generation.

Yep.

And one of the dark sides of AI is that it's so easy to go 0 to 1. that you might just be doing it too much and that picking the right battles and choosing what to build, what to generate, and maintaining a high level of quality is going to become a differentiation. And I ultimately think quality begets trust. What is trust, especially with a software business and software company like Salesforce or Vercel?

Yeah.

Just quality of customer obsession and service proven over time. That's what I think trust is. To me, Ratatouille embodies that. Pixar embodies that. And also, it has this hilarious part of the movie or theme, which is this saying that anybody can cook and encouraging people to try things and build things.

Yeah.

And I think AI is helping us all do that, right? Obviously, talking about my book, because Vercel helps everybody within the organization build applications and deploy them. securely. But I truly believe that AI is giving us all superpowers, and that is truly a killer app.

Yeah, let's talk a little bit about quality. You know, I just hired someone to take my job running Slack engineering, but when I was running Slack engineering, you gave us a very hard time. You had a lot of very specific feedback on very detailed things about Because you use Slack, you use Slack heavily, you're a big customer.

We run everything on Slack.

But you had a lot of detailed feedback, and I would equate it related to quality, your expectation, your bar for quality. I think Slack has a lot of quality, so it's a very high-quality product, but your bar was very high. So talk to me a little bit about that, and your expectation of speed also.

Yeah.

Like, you were like, why isn't it fixed? Why haven't we done this? We have to go fast. So maybe give me a little aspect of quality and speed.

As a spokesperson for a lot of engineers in the world, a question that I get very frequently is, as an engineer, am I going to have a job in 5 years, in 10 years, in 15 years? Creating high-quality software is extremely tedious and expensive.

Yeah.

I actually give Slack as a great example of something on the surface seems like it should be easy to generate something like Slack because you look at it, it's like, oh, it's a sidebar. channels, conversations, how hard can that be? But the depth of quality of engineering around security, around scale.

Yeah.

I remember when you guys, Slack, went vertical during COVID right? Just supporting the throughput of messages per second. And so there's all these dimensions of quality, scale, security, reliability, And what we try to do at Vercel is your next incremental piece of software has to always be higher quality than the previous one. We've been excited talking a lot about AI software factories because as code becomes free, the value of, or the differentiation of a firm shifts one layer up. into the best practices, the institutional knowledge, the context engineering that you put into your factory to produce that next incremental piece of software. For example, something that I care deeply about that you just called out is speed. And over the years, I collected a lot of nuggets of wisdom about speed. Speed is not just, I measure a transaction in number of milliseconds. Speed can also be the perception of speed.

Yeah.

A great example, I give you guys a lot of feedback because I travel so much. I'm on a plane all the time. Occasionally, a plane will be Starlink-enabled and my happiness level goes through the roof. More often than not, it's Gogo Inflight or whatever. And what happens with these internet connections is that they drop a lot of packets. If they drop a lot of packets, it makes messaging your friends, family, and coworkers really, really hard. And so what do you do in those circumstances? You have to provide a pleasant experience to the user, even though the network is unreliable and slow. There's things you can do to make that experience more pleasant. Now, I've been studying real-time systems and human-computer interaction for decades. Now, how do I transfer that to the next generation? How do I enable my coworkers to use AI plus these best practices. So, we work a lot on codifying this into skills, into frameworks, and into tools that make this factory work better.

Yeah. Well, it's kind of, in our keynote, we talked about determinism and probabilism, hard to pronounce. Salesforce is definitely, you know, really working on providing frameworks around the right context at the data layer, the right application definitions of your business processes, your semantic layer, the right definitions of AI. Because anyone can vibe good at anything. These are new interfaces.

Yeah.

AI forest layer, as we call it. It can look amazing, like, wow, that's incredible. But it might be the wrong data. It might not be secure. It may not even be a very good interface. So how do you think about allowing— like, the power of building is now in everyone's hands. You're helping make that happen at Vercel, which is incredible. But how do you think about giving that power but adding those guardrails to guide to security, to quality of user experience, to all the things that I would call trust?

Yep. Something that has created or enhanced the trust that our customers have for Vercel is our care for design.

Yeah.

We care a lot about the craft of our interfaces. We care a lot about the consistency of our interfaces. And so, if everybody can cook, can they actually all design beautiful interfaces? Well, we run the experiment and the answer is no. You ask an agent build me a website or build me a sales deck, and they all look the same. It started out with purple gradients.

And it doesn't look like Vercel.

It doesn't look like Vercel. Started out with purple gradients, then it became beige stuff with serif fonts. And you look at it and it's like, well, it's really cool that we can do this. By the way, I don't dismiss that at all. It's amazing that anyone can transform prompt into artifact. But how can you infuse the brand and design principles of your company? So the first thing we did is we took this standard that Google pioneered called Design.md. It's basically a skill for design that brings some more determinism.

Mm-hmm.

So I really like that framework that you just put out, combining the good parts of AI, which is the creative stochastic nature of it, with some determinism. So Design.md, for example, can codify your color system. Obviously, we like black at Vercel, but we like some other colors and we listed them out. And then you can actually use some more stochastic stuff. You can write down in English, hey, if I need to produce a report for sales, avoid this and avoid that and use my tone of voice here and there. So design.md helped us a lot, but we took the deterministic guardrails even further. Yesterday, we announced a linter for enforcing design system constraints.

Mm.

So one of the— and this is just an example of the cool things that you can do, but I think the future is going to be combining these 2 things. I call it verifiers and skills. Skills alone are just guidelines. And just like humans can sometimes mess up and not follow guidelines, You need a hard verification system. You need something to enforce constraints. In engineering, there's a programming language, Rust, that is growing in popularity because Rust imposes very harsh constraints around memory safety. And so combining the loose of AI, the loose vibes of AI with hard, almost like mathematical proofs, is, I think, the future of a lot of this, where software engineering is headed.

Yeah. Have you heard about the SaaSpocalypse?

I have.

Well, we've been talking about that, that it's not really true, but I've heard you talk about, you're not going to vibe code all of your systems in the enterprise.

100%. So first of all, I think the discipline to decide where you put your tokens, So I think the first wave of AI was CEOs, and for the record, I didn't do this, but CEOs were holding leaderboards of like, please spend as many tokens as you can.

Token maxing.

Token maxing. I always thought it was the dumbest thing ever because fundamentally tokens are compute. Would you ever wake up one day and say, hey everyone, go to the AWS console and create as many instances as you can? fire up as many computers as you can. Well, I would fire that engineer saying my cogs are going to go to the floor, right? And so with tokens, it's the same. It's just GPU compute.

But we were for a moment comparing token usage to productivity.

Yes. I think that's wrong. And I also think, again, the question that we need to answer is there's some finite amount of tokens that you're going to use, right? Where do you put those tokens towards? Is it towards hardening your cybersecurity? Great use case. We launched DeepSec, a cybersecurity hardness. I believe that that's one of the best things you can do with AI today is strengthening the reliability and security of your systems. Quality assurance. Agents with computer use and browser use can literally navigate your interfaces and your programs and tell you where you're messing up. So all of the things that sort of increase quality, I think, are great uses of AI. Now going to SaaS apocalypse, I think what AI is showing is that systems of record are extremely valuable, but they have to have the right agent ergonomics. So agent ergonomics is the— or agent experience is the term that I use for if I have a bunch of data in my CRM, can an agent access that data? Do I have the MCP for it to access that data? Do I have the CLI, the SDK? Is the data fast? Is it scalable? One of the things that we're going to have to confront is that agents will use compute far more than human beings. I'm a heavy user of Slack, but let's say I send 1,000 messages per day, right? Or maybe that's too much, but I send a handful of messages per day. Just imagine how much more agents can do because they never get tired. And so they're going to query your systems of record like you've never seen before. I think you guys were sharing some data on this. You launch an MCP and all of a sudden the demand for it seems nearly infinite. And so I think the SaaS apocalypse is— what it is about is customers are going to generate a lot more software. They're going to access a lot more data.

Yeah.

And those repositories of data are going to become extremely important and they have to have the right interfaces for agents to be able to use them.

So let's talk about interfaces. You know, we call it headless. That's our strategy.

That's right.

And Claude Force with your AI Force is the overarching brand. Maybe launch Claude Force. You saw that and the amazing artifacts. Those are incredible. Slack calls them surfaces. So in the keynote, you saw that. How do you think about Vercel and Salesforce? 1+ 1= 4. And what are use cases?

1+ 1= infinity, right? Because When I started at Vercel, I realized very quickly that the value that we're bringing to the world was in the customer experience side of things. So a lot of our bigger customers were saying— great example, customers like Under Armour and Nintendo, they wanted to innovate on the customer experience side, but they wanted to keep using Salesforce as their source of truth and customer data repository. And so they combined Salesforce and Vercel, And they launched amazing web applications and websites way faster than they could have done if they had rebuilt these things independently from scratch.

Yeah.

And so that's why the 1+ 1= infinity. Now, the limiting factor back then was if you wanted to move fast and create the next underarmor.com, you needed a DevOps team. Vercel solved that. You needed to create frameworks like user interface frameworks, et cetera. Vercel solved that with Next.js. The other limiting factor was writing code. And now that's not a limiting factor anymore. We can use coding agents to generate these amazing interfaces, to launch new products. And so, like I mentioned, I think this is a match made in heaven. Also, Slack Code, I mentioned I run the company on Slack, Slack Code for me was the missing piece because so much of improving a product emerges from conversation. Conversation plus customer data plus feedback that we get, and then we digest it together. The best software has always been a byproduct of human collaboration. So now it's human plus agent collaboration in one place. So really bullish on where things are headed there.

Yeah, glad you like that. So for those of you who don't know, Slack Code is, it sounds like coding in Slack and it is, but Slack's not offering the coding. So partners like Vercel can actually use the interface of Slack for multiplayer coding and get all the value of Vercel plus the value of Slack.

And the value of, again, like the human oversight is still really important, right? One of the most popular words now in AI is steering. Because what happens is the agents are working for longer. It used to be that, I mean, it seems like forever ago, but just a few years ago, all AI for coding was, was autocompletion.

Yeah.

Copilot inside your code editor, you start typing, it autocompletes, and our minds were blown. Later we got ChatGPT and we could ask it and you could build something and you would copy paste the code. Now agents can work on code for hours. And that happened in just a few years, but they still need oversight. They still need steering. And I actually find that the Slack interface is great for this because you and I can, we might be debugging a customer escalation. We might be debugging an incident like Resolve.ai just probably spoke about. We could be brainstorming a new capability. And so I think that steering of the agents by the humans is still very important. I don't want to live in a world where it's just the agents and we're not participating, right?

So how do you think we need to evolve as humans and our skill sets to be better at working with AI, at delivering that quality, the trust, without breaking trust?

There's a term that I've been using called agentic inquiry. A big debate in my engineering circles is, should we read the code or should we not read the code? Or if we're reading the code, for how much longer? And my answer was like, I don't know.

How much longer will you be able to?

Yeah, exactly. Are we going to read the code in 3 years or are we going to treat it as assembly? Are we going to treat it as zeros and ones? And my answer to that is we have to understand the systems. That might not mean necessarily reading the code, but we can ask the agent questions of the code. We can have the agent work for a few hours or minutes or whatever and say, hey, show me what the new architecture looks like, or run a simulation. Or— and a big part of what Vercel is proposing to the market is don't get locked into one model. Have the models debate one another.

Yeah.

Imagine being in a coding channel and Vercel is there, And Claude Code is there and Codex are there. And let's have this conversation, right? And Fable critiques SOL and SOL delegates to DeepSeek to save on costs. And so I see a world of multimodality.

Yeah.

And the humans need to understand we do need to change, I think, some of our expectations. We need to reset and update our expectations very frequently.

Yeah.

I think if you get too stuck in your old ways and saying like, well, I'm sure agents can't do that.

Yeah.

That's a big mistake. Agents can do pretty much everything with the right guidance, steering, and skills. The other thing that I've learned that a lot of people are doing at Vercel is asking the agent what questions we should be asking them. So the idea of prompting presupposes that we know the questions. So sometimes, You can lean on the agents to say, hey agents, if you were CEO of Vercel, what should I be paying attention to?

Mm-hmm.

We have an internal agent at Vercel called V that is connected to everything inside of our systems, Salesforce, Slack, our customer insights, customer feedback, X. And we can ask this system, hey, what are the trends this week? What should I be paying attention to? What are interesting things that people are chatting up about in Slack? And so I think trusting that the agents sometimes know better than us can be a challenging thing to do.

All right, this is a bit of a different question, but back to the human element. Do you think people should still study computer science in college? Or what disciplines do you think are going to be most valuable in the future, if you were to tell young people going to college right now what they should focus on to prepare for the world that we see now and as well as what's coming, what would you say?

You know, I was doing this chat with Paul Druce from Salesforce Ventures earlier, and he was asking me about my early days of— I grew up in Argentina, I taught myself how to code, And I remember the things that I would find exhilarating, like compiling a program and getting it to work. I was like 10 years old. I was like, oh my God, this is the best thing ever. And I think about my kids, my oldest one is 10. Compiling a program is like no problem. Even deploying an application on Vercel that could scale to a billion users is literally one prompt.

Yeah.

When you deploy to Vercel, the infrastructure is so well tuned and so scalable that my kid could prompt and ship an application that could reach billions of people. That was unheard of when I was 10. That means that what I want for him is to now raise his ambition by 100 times. Okay, you have these amazing technologies. What can you do with them?

You got to get those billion users.

You need to get 10 billion agent customers, or I don't know what I'm going to set as his bar, but Yeah. I think he can go further in creativity. He can produce video assets. He can design. He can do so many more things. And so I think the future of computer science will be harnessing all this power. Computer science in the past was punching cards, then compilers, then getting programs to run, then publishing websites on the internet. So I think we keep making those things simpler. seem obvious and easy, but then there is another big next bar to accomplish.

All right. A lot of talk lately about cyber and the models and, you know, or even bioterrorism. What should we provide as vendors? I'm not going to ask you your opinion of where the world's going, but I think we all— I really liked Jensen's, I think it was Jensen today, is like, you know, don't ship it if it's, you know, and come up with the right testing. But what more can we do as software vendors to provide the tools for the future, for the risks of the future?

Yeah. One is we published an essay called Everything Hackable Will Be Hacked. It's very important that people—

That's positive.

Yeah, I guess there's a positive angle to it. So, we really needed to get this message out into the world. The agents of today, Dario is saying about the dangers that come in the future. I'm talking about the dangers of today are far more real and far more serious than we comprehend, I think, in a lot of cases. And so the ability for coding agents today to hack into systems far exceeds what anybody is talking about. And so what we did is we created a benchmark for cybersecurity and we started testing models and we started realizing that there's way more vulnerabilities than we can even have patches for. And so the message that we deliver to the world is start putting your tokens towards defending your systems. Start running the scans, start patching the bugs, because they're already there. We're going to be putting more money and more resources towards hardening the open-source infrastructure of the world.

Yeah.

So Linux, I'm a huge Linux fan. Linux changed my life. I was a kid and I couldn't get Windows to be a good developer platform. And I switched to Linux when I was 11. And I love Linux, but we have found a lot of vulnerabilities. It was code written in C. It's not a memory-safe language. And it was written by humans. So it has a lot of— humans make mistakes. And so we're going to have to upgrade a lot of our global infrastructure. And again, it's not about the science fiction Skynet things of the future.

Yeah.

It's the problems of today. And so that is a thing that we have to do ASAP. The other thing is what's fascinating about agents is that they get better the more power and the more data you give them. So when ChatGPT came out, it was really awesome, but people were saying, can you write me a haiku about Salesforce? And we're like, oh my God, it wrote a haiku. And then we realized, well, what if it has more context? Oh my God, it gets more intelligent. And then we realized, what if it gets more tools? Oh my God, it gets more intelligent. And then we realized, what if it gets a computer with an internet connection? Oh my God, it's even better. And so agent IQ grows commensurate to the power and resources that we give it.

Yeah.

And therein lies the problem. If we don't sandbox these things and if we don't govern how they access data, then they will create risks and they will exfiltrate our data and they will hack Hugging Face. And combine that with the fact that they're being trained for running for longer, And supervised, which again can be a great thing. You tell the agent, hey, solve the Navier-Stokes equations. I'm going to go to the gym. When I come back, please solve that Millennium Math Problem. It can do that, but it can also do things that you don't find positive or that you expect to happen. And so sandboxing, we created the Vercel Sandbox, which is a way that you can build agents and sort of control what they can do. You can put a firewall around them. And so this requires a lot of innovation, a lot of new infrastructure that doesn't exist yet, but we're working on it.

Well, it's interesting you said Linux was created by humans, so there could be a lot of mistakes. And then you didn't say this, but it implies that AI does not make mistakes.

Oh yeah, it makes lots of mistakes.

Clearly an AI was trained on all of this.

AI makes a lot of mistakes, but I do want to say something. positive about the AIs because you can combine them with these verifiers. A colleague of mine called it proof engineering. So I don't know if you saw the repository for the Navier-Stokes proof that OpenAI put out. None of us can understand it. They use a proof checker called Lean and they open-sourced it. Kudos to them. And then you can go and see all of the proofs that this thing wrote. It's basically millions and millions of lines of mathematical proofs, but they work because we have a compiler, we have a proof checker called Lean, and so we know that it's mathematically sound. Very few people in the world can understand what it even means, but it works, right?

Yeah.

And so this is a risk. I think on one hand, AI can create provably correct systems. doing things that we don't fully understand. It could be arriving to results that we don't fully understand. And so some of that is scary, but again, we can put guardrails around them and we must.

So you can build it into the software factory to say people can go faster, but those things will be built in to—

Maybe to make it super practical for the software factory, you can put guardrails around performance.

Yeah.

You can tell the agent, look, this critical transaction, maybe for Salesforce, their MCP query that says, Look up customer by name. I want to have a P99 of 200 milliseconds. The agent will go to whatever extent it needs to go to satisfy 200 milliseconds. And it's amazing to watch. Now, the same happens with cyberattacks. In our article about everything hackable will get hacked, we started reading the reasoning traces of this open-weight model called KIMI. And when our CTO started reading what the model was thinking, he frankly got a little bit scared because when you give the agent the task of find the vulnerability, it's so resourceful, it's so unrelenting. Human hackers get tired and they need to get paid and they need to eat. These agents, just you give them an objective and they go, go, go, go, go.

Yeah.

So it can be a positive and it can also be a negative.

All right, my last question. Um, so when I was working with, in the Slack unit, one of our best developers who just totally, it was actually through his child, it was kind of funny. So Claude 4 or 6 came out February, I think, like kind of hit that tipping point. And he had tried previous models, versions, and, and he's a very, very good engineer, wasn't that impressed. And then, um, He was working with his daughter and he said, well, let's build something. And he's using Claude 4.6. All of a sudden he's like, wow, this, the world just changed. It just blew his mind. He's one of our most productive engineers and he doesn't just use one agent. You know, he has an agent, you know, doing the plan, doing the spec, another one debating. He's got one writing the code.

An ensemble of agents.

It's like an orchestration of agents. You know, in a negative way, the Hugging Face incident, It was a, you know, a community of agents.

Yeah.

I think they called it, um, working together. He, in a very small way, has a community of agents, but he's kind of doing it in a bespoke way. If I take kind of the middle developer in our stack rank of really good developers, they're producing not as good code. It's not, not as fast, or he or she, just use it as an example. And hasn't really learned the pattern that the best developer has. How do you see the future where we can, instead of saying, gee, you need to do it like this guy, how do you bake it in so that everyone can go as fast as our top developer?

Yeah, for sure. So I also have arrived to the same conclusion, by the way, that the best agentic engineers are Harnessing multiple models. It's not just multiple models, it's multiple roles. And multiple roles within them, absolutely. And by the way, this will save you costs. You'll be really happy to hear, right? So with AI Gateway, we offer developers one API key, multiple models. And one thing that we've been noticing is that the OpenWeight models in aggregate token use are now towering over how much the proprietary models are used. And so when we're designing these pipelines for engineering, we have to think about using models and using agents much like systems, right? This is why I think the factory analogy is cool to think about because you're designing this sort of production line.

Yeah.

And so at the beginning, you have an agent that sort of creates the blueprint. And you might use a very, very smart model for planning. And then that hands it off to another part of the factory that swarms out to implementation. And then every factory has a quality assurance step, right? And you have a different agent in a different role using different tools and being given different instructions.

Yeah.

So it's mesmerizing to watch agents these days do QA, by taking screenshots, looking at them and saying, yeah, this looks good, or running benchmarks.

Yeah. And it's crazy. The AI is looking at the production code.

And it looks at the results.

Looking at the UI.

It looks good. It reflects to itself. And so going back to what do we bring to the table? Well, I think it's designing these pipelines. I think the best practices of engineering today are exactly what that engineering is doing. thinking one layer above, how do I give my job to the AIs?

Yeah.

And the other one that I think is really relevant to this audience is at the end of the day, we have to make customers happy. We're all working in service of customers. So infusing this factory with your context of, you know, you had a call with a client and they give you a bunch of feedback, they're filing support tickets, Can you automate the— can you feed that into your factory so that it's not up to— even a really good engineer might not just have the context of what to build. And so helping them prioritize, I think automating this software development lifecycle, we call it the agentic software development lifecycle, is the future of software.

Awesome. All right, everyone, please thank Guillermo Roig, CEO of Vercel, founder. I hope you enjoyed this. Enjoy the rest of Dreamforce. We got a lot more coming at you, but thank you for being here today.