Interview: Paul Klein (Browserbase) on how AI agents using the web is unlocking new automation

Foreign.

Guys, we're back here at Re Invent 2025. I'm here with Paul from Browser Base. Really excited to have him stop by the work OS booth. Say hi Paul. Really quick, introduce yourself. And what is Browser Base for people that don't know?

Yeah, of course. Hey everybody, I'm Paul. I'm the founder of Browserbase. So we're building Browser Base. It's a base for your browser to see the name. But what that means is we provide web browsing capabilities for AI agents and applications. You can think about AI agents or AI models as being trained in all the knowledge you've seen on the Internet so far. But what if you want to build an application that does more of that knowledge, that actually takes actions on your behalf? A lot of the work that you and I do happens in a web browser. If we want AI to do work on our behalf, we need to give the tools that we use every day, and that tool is a web browser. So browser builds the infrastructure, which is a browser infrastructure. We run browsers in the cloud. We have an open source framework called stagehand that enables developers to control a browser with code. And we also have a product called Director AI, which is almost like a low code way to generate browser automation. So when you think about automating browsers or letting AI do actions on your behalf on the web, browsers kind of powers a lot of that stat for some large customers like Perplexity, Vercel, Clay and Kimura.

Cool. Awesome. Why is this hard to do? So I think in the past people probably know they've run headless browsers for like web testing or integration tests. Why is this sort of a different problem? And why is it a hard thing that Browser Base is able to solve for customers?

You know, like the, the idea of a browser, like using a browser in a cloud environment, it's kind of a hack in itself, right? Browsers designed to run on desktop hardware. You and I, when we use browsers, we probably have like a hundred tabs open. It's using a lot of our CPU memory. So doing that in a lightweight cloud environment securely is really challenging. Browsers almost like little nuclear bombs, right? They could go to any website in the world and that website could exploit, you know, something and download a malicious file. And do you want that in your AWS cluster? You want to make sure that your browser is running performantly, it's secure, and you also want to be able to look at it. You want to see the observability, you want to see where the browser went. So we think that building production grade browser infrastructure so you can know where AI goes, you can enable it to do that successfully and you can make sure it doesn't do something you don't want it to do. That's a really hard challenge and it's completely what we're focused on at browser base.

You mentioned some of the customers that work with browser based today like Vercel or Perplexity. Can you walk through how they use it? Like what, what are they trying to accomplish by using browser based browser actions?

You know, generally browser based powers, so many different use cases and you can think of the, the hyperscaler, the big companies, but I also think about the really really small companies. There's a 50 year old dairy company that's a browser based customer. They had never hired an Engineer in their 50 years of being a company until this year they started playing with AI and the first thing they wanted to do was they had Nancy in the back office looking up gas prices every single day to tell the truckers where to stop for gas. Instead they built an AI agent that would go to Google Maps, search the route, look for gas prices using the browser and aggregate which place they should have the trucker stop for gas. Super simple use case. They spun it up with Browserbase and now Nancy can spend more time on interesting tasks as opposed to having to go look at the gas prices every morning. And this kind of work carries over to so many different areas like Kyb. If someone signs up for a bank account, how do you make sure that you get more about that person outside of what they're putting on the web form? How do you go to their website for a business bank account if you're a founder, Pretty hard to bank you if you don't know how much money you've raised, you might want to go check Crunchbase or Pitchbook. These are places where you have to use webcasts. One of our customers, Benny is really cool. They they do rebates for food stamps. You might not know this but for food stamps or ebt, if you buy healthy food, we're going to US government will give you a discount. They want to encourage you to buy healthy food. But to get that rebate you have to submit this web form with your receipt and it takes like 10 minutes and it sucks. Benny is a consumer app where you take a photo of your receipt, they go to the web form on your behalf, upload all the information and you get a rebate for your food stamp usage, which I think is really cool. We've done hundreds of years of browsing this year at Browser Base. So we've really like deleted hundreds of years of people going to websites and clicking buttons from the universe. And that's really exciting about all the different types of work you can do with the browser. It's meaningful time save for our customers.

As you're describing this, it kind of reminds me of people previously building like web scraping which was always like pretty brittle, you know, it was like unreliable. But now with AI you're able to bring this like more I guess like dynamic interactive task based approach to it. Is that right? Are you using, is this using like the AI models under the hood for the browsing?

Yeah. So like web scraping or rpa. Robotic process automation are like the old kind of categories here. Old acronyms, old acronyms. But maybe AI agents are just RPA plus plus is what some people are saying. Right. But really what we've done now with Browser Base and really our framework stagehand is you bring the model. We don't publish any models, but we help shape the context and provide the interfaces for the AI models to interact with the browser. So what that means is that you might have a natural language command, click this button and when you go from the browser to the AI model, you need to expose the right tools in the browser that AI model can control. Now an AI model might just move the mouse to certain coordinates. It might need to like get the whole source code of the page, but that's pretty complicated. Um, instead you can expose simple tool calls like take this action, extract this data, navigate to this page to a model, give it more kind of a condensed context window to actually execute those commands. So with browserbase we don't bring any of the models, but our customers use their models with our framework statechan to actually build production grade web automations that don't break when a webpage changes. They keep working and self feeling on their own.

It's pretty cool that people can bring their own model to this. Or you're building something that's sort of like a scaffold around the models that are out there. Do you, is there a model that performs the best that you found for this or your, your personal favorite?

Yeah, so I mean, I mean it's funny. I'll say this answer and I guarantee you in like two or three weeks it won't be the right answer anymore. I can't take a vacation. Models keep coming out. We actually published this whole evals website called stagehand.devevals that will show us every Single like the performance of models on like common things we've seen. For example, at browserbase, we've browsed through, you know, millions and millions of websites. We've seen so many different types of date pickers. So we can recreate all those date pickers and see how effective a model is at navigating a complex date picker. There are so many ways to specify date on a website. We've seen thousands of them. So in our emails recently, Gemini 3 is still a really good model. Before that, 2.5 Pro was the best model. So we think that Gemini line of models have done a really fantastic job. We actually worked with gemini in the DeepMind team on their Gemini computer use model. They launched benchmarking that ahead of the announcement to kind of show where they were and they were state of the art when we published it. So we're very big fans of Gemini. We also just worked with Microsoft on their recent Fara 7B model. Now the Fara 7B model is very interesting. It's an open source model, open source computer use model that's a fine tuned version of a Gwen model and it's very small but it has, you know, comparably good scores relative to other models of its size. And being open source, you can then run that in your own environment. And I think customers at work OS who are very authentication or security minded, they probably care a lot about model security, AI safety and a lot of them are probably running their own models because they want to make sure they own the data that's going through this type of model. I'm not sure if you've seen that on your side of the house in the business.

Yeah, I mean I've seen examples of people taking the core models and fine tuning them or you know, using it as a foundation to like jump off to build their own stuff. Pretty exciting. You talked about Stagehand, which is open source.

Yes.

So tell, tell folks more about that. How did it become open source? What's open source? Like, what's open source? What's not open source? What's the relationship with Browser Base? How is this part of your product strategy and ecosystem?

You know, I think you can. Our business strategy is comparable to other open source companies like maybe a Vercel, where you have Next JS being one of the leading open source frameworks for building websites. And then you have Vercel, one of the leading ways to run those websites in the cloud environment. Right. Browser Base has stagehand, an open source framework for allowing you to automate a browser. You can Imagine it as a Playwright, puppeteer, Selenium, if you heard of those frameworks, but powered by AI and browser based is the infrastructure to run those browsers connected to you by Stagehand. So really we have three primitives in the open source framework. One's called act, one's called Extract, one's called Observe. And the kind of interesting part of Stagehand is that it's a framework that uses natural language as the input. You aren't saying, click the fifth button on the page, it's red. You're saying, click the login button. And an LLM then takes that instruction and compiles that into browser commands behind the scenes and caches those commands. So if the login button changes to sign in, it will still be able to reprocess that and find the right button, even if the page changes. Now that's completely open source MIT license. And you might be asking, like, why would we open source this kind of important part of our business, this framework layer? Well, our belief is that we need to do whatever we can to make browser adoption easier for people building AI agents. And when we first started the company, it turns out that the framework layer really wasn't built to kind of satisfy the demands of AI agents. People have built these browser frameworks like Puppeteer, Playwright and Selenium for testing, not for automation and not for AI. So Stagehand was our answer to building an AI first browser automation framework that really worked with the models natively as opposed to being an afterthought. And that really accelerates the adoption of our browsers because once again, our north side metric is how many years of time are we saving people every single month? And we want to keep that going up. And that means making it easier for everyone to adopt these browsers, um, by any means necessary.

It's awesome. Um, what are you seeing in the enterprise in terms of adoption, I wonder? You mentioned, like, some of the growth stage startups like the Perplexity Vercels, and then a lot of kind of smaller organizations like the Dairy Farm, you know, like, like organizations where they may be adopting this technology for the first time. Are you seeing any pull from bigger organizations?

Dairy Farm is a $50 million business here, by the way.

Okay.

Yeah, don't, don't hate on Team Dairy. All right. I drink a lot of milk. What was the Enterprise? Yeah, I think the Enterprise is a really interesting story right now. I actually just spent this time at what someone's called the Davos of CIOs meetup, which in Half Moon Bay is Run by Wing Ventures and self appointed Davos, by the way, how do you name it?

That secret cabal of CIOs. Yes.

And they put up this chart which is really interesting, like what are people using AI for in these organizations? And the number one very obviously was customer support and number two was software engineering coding. But a lot of the stuff that I think is really interesting and where we see in browser base, which is like finops automation, back office compliance or rpa. Interesting. Procurement automation you can do that stuff is still pretty early actually. I think the browser automation and computer use space is a kind of like where coding agents were a year and a half or two years ago. And that's mostly because model quality, but also just awareness like coding is a very clear problem. Engineers are at the forefront of AI because they want to automate as much of the work as possible. Engineers have always been inherently lazy. So for enterprise adoption of computer use I think it's still early, but we're poised to see a huge investment. As far as I understand it, every single AI lab is investing so much money into training and developing computer use models right now. So my prediction for 2026 is that you'll see step function increases in computer use capabilities, that becoming a first party model offering for every AI lab and downstream effects work workflows are use cases that were really challenging to build one year ago, become very, very straightforward and easy in 2026. And of course I'll trick down the enterprise where enterprises might be able to now self host models that have, you know, progressive improvements. Bar 7B is a good example of an open source model. You can sell those. The infrastructure is getting better. Browserbase plays a key role here by providing great infrastructure where you can go to any website reliably every single time. That's important because a lot of times the models are good enough, but the browser infrastructure isn't. We need to make sure the browser infrastructure is just as good as the model capabilities. If those three things kind of come together in harmony, I think we'll see huge explosive growth in this space in the next year. And I'm frankly super excited for it.

Yeah, unlocking the web for agents and AI systems. It still is this thing where even despite the enormous growth we've had in AI, it's still not in that layer of the technology stack.

We've all used OpenAI operator and it kind of felt pretty mid, you know, and it was exciting, it was a big unlock of what can happen. And I think there's two kind of things that govern my business's growth. One is model quality and the other one is inspiration. I'm doing everything possible to partner with the labs and we just did that with Microsoft and Google DeepMind on evaluating the models to improve model quality for computer use. But the second part is inspiration. How do I show people what's possible? How do I show them that every single time you have to do a routine task, an AI agent can do that for you if they have the right tools and models? I spend a lot more time on inspiration these days. I think people are going to be very inspired when they see what our customers can build and do. And we'll be publishing a lot more case stories in the next 12 months about all these interesting things people are building with browserbase.

Awesome. Love it. The last question for you, we are here at Re invent the craziness that is the giants pretty crazy conference. What are you most excited about that you're seeing here this week?

Other than work os? Other than Michael, of course, other than work os, but other companies, other products.

Other things people are building. What's your week been like?

You know, I constantly have been talking to enterprise buyers of AI and ask them how's it going, how are things going? What's their attitude towards it? And I was really, I was really scared that people would come in and say, well, it's a bubble, this AI thing's a bubble and we aren't going to invest in it, we don't care about it. But people are very excited and optimistic. There's enterprise buyers that are wanting to figure out their roadmap for 2026. They're excited about AI at Reinvent. And Reinvent kind of often lags behind the, you know, fast paced Silicon Valley stuff that we're used to on Twitter. So I'm really excited to see this attitude of optimism, this attitude of what can we do or what should we try and not what shouldn't we try and what are we afraid of? There's a lack of fear and a huge amount of opportunity excitement from enterprise buyers and that's where you really see these at scale deployments of AI systems and that's really optimistic for me.