Why AI Agents Need a New Kind of Browser  ## Metadata - Author: [[The New Stack Podcast]] - Full Title: Why AI Agents Need a New Kind of Browser - Category: #podcasts - URL: https://share.snipd.com/episode/51db0909-7ffc-4aa9-9f97-97663935ccbf ## Document Notes ## Highlights - **What is a Headless Browser?** - Headless browsers are real browsers running without a visible interface, crucial for web testing and automation. - They're sophisticated software environments that render UI, execute JavaScript, and simulate human interaction programmatically. Transcript: Paul Klein Yeah. Yeah. Well, there's one thing I know well, it's headless browsers. And maybe defining that term is important. It's one of the weirdest names in infrastructure. Our type of infrastructure is called a headless browser, like a headless horseman, right? It was a little silly. But we could even go all the way back to Netscape, right? The origin of the browser is so important. And what does a browser do? Well, it takes code and renders user interfaces, and then allows for people to interact with those user interfaces. And, you know, the browser really built the distribution engine on top of the internet to allow very useful software to have a universal interface where everybody can access software Online using a browser and a set of standard programming languages, HTML, CSS, and JavaScript for the most part. Now, as developers built more software that exists online, more websites and web applications, they needed a way to test those applications. Because it turns out it's quite tedious after shipping a feature to go into your new feature on a website and click all the buttons 100 times and to continue to do that every time you change A feature. So developers started thinking about ways that, okay, now we have people using browsers, but for our software testing workflows, we need to have computers, or CICD, is kind of continuous Integration, continuous delivery of software, needs to be able to test our software. So they created, well, let's make a browser that can be controlled by some code. And this code was very static. It was saying, go to this specific button, click this button, and it ran in a way where it was headless. Nobody needed to see this browser, so it kind of just ran in the command line. It still rendered all the UI elements. It still ran all the JavaScript, but it didn't actually show a display in a way that you would see your normal browser. That's why it's headless. And these headless browsers were really useful for testing. And a ton of companies have built great products around headless browser and simulation testing for websites. I was doing this in Jenkins way back in the day. We were running browsers in Jenkins with headless browsers with PhantomJS or Cypress. There's BrowserStack, a very large company that does a lot of headless browsing for testing. SauceLabs has done a ton of great stuff here. And there's always been innovation in the testing side, and several frameworks have come out to make it easier for developers to integrate browser control into their tests. Those frameworks are called Selenium, Puppeteer, and Playerink. Now, the important thing to note, they've always been built for testing. And this space has always been very testing-oriented. Now, BrowserBase isn't a testing company, foreshadowing here. And that's something that we thought about. If browsers are useful for testing, why haven't they been used for more automation? And what does change? So the world so far was mostly testing. Now, what's changed? And it makes sense for testing because if the test breaks, you want to know. If the button changes color or changes text or changes location, your code breaks. These tests are tightly coupled to the interfaces that they're automating. But if you want to automate something on behalf of people, if you want to automate actual work, you need to have more durability. And the frameworks and companies that have existed so far and the infrastructure has been really oriented around being quite brittle by design so that if your website breaks, you know About it and can fix the website breaking. But if I'm building software that's automating work on behalf of a person, if the website changes, I still want my automation to continue working. And that's really introduced a lot of opportunity for building browser automation for the future, which is for actual like automation of work, not for the existing world, which is Mostly testing. And that's where browser-based has come in. And our framework Stagehand, which really kind of is a superset of the existing frameworks, Puppeteer, Playwright, and Selenium, it's actually much more durable than those frameworks Because it does a lot more fuzziness. And the kind of the advantage here is that it has an LLM functionality, an AI functionality built into the framework. In the old world, you can say, click the login button. It's the fifth button on the page. The color is red. It says login on it. In the new world, you can tell AI, hey, I want to click the login button. You figure that out for me. And if the login button changes color, maybe changes position, maybe changes from login to sign in, you can still find the button on the page using large language models. Large language models are very useful at parsing code. So as we think about what's possible now, we can now build much more dynamic functionality with headless browsers because we have LLMs that can on the fly figure out what the appropriate Action is to take on the page. And that's really increased the size of this market. In the old world, if you wanted to automate 100 websites, you have to write 100 scripts. In the new world with AI, you can write one script that can control hundreds or thousands or millions of websites by using these low-level dynamic primitives where LLMs can look at a Page and take that next intelligent action. That meets a demand for more browsers. That's why our infrastructure platform exists. We're here to help cutting-edge startups power their web browsing capabilities of their AI agents or AI-powered software by being the infrastructure layer, the tool call layer, And then also an application layer for really building your first POC. Frédéric Lardinois If you don't mind, Alex, let me just jump in there for one second. In terms of how you're actually working with the browser, because there was a time where it felt like, okay, maybe we have to go into the DOM, right? Like really low level to interact with the browser. How are you doing that now? Are you using a real browser? Like what does the headless browser actually look like for you? Paul Klein Yeah, a headless browser is a real browser. So it's the same, you know, we use Chromium and open source browser. And you need a real browser because what is a browser? Well, a browser is a very sophisticated networking engine. It makes a lot of web requests. It's a very sophisticated rendering engine. It renders HTML and CSS into UIs. And it's also an interaction layer. When I click a button, it's a JavaScript runtime. ([Time 0:03:56](https://share.snipd.com/snip/2af9548a-65d9-4da3-b9de-6ee30ad4c423)) - **AI Makes Browsers Resilient** - Traditional headless browsers are brittle and break easily when web elements change. - AI-powered scripts can handle thousands of websites by dynamically interpreting actions using large language models (LLMs). Transcript: Paul Klein Now, what's changed? And it makes sense for testing because if the test breaks, you want to know. If the button changes color or changes text or changes location, your code breaks. These tests are tightly coupled to the interfaces that they're automating. But if you want to automate something on behalf of people, if you want to automate actual work, you need to have more durability. And the frameworks and companies that have existed so far and the infrastructure has been really oriented around being quite brittle by design so that if your website breaks, you know About it and can fix the website breaking. But if I'm building software that's automating work on behalf of a person, if the website changes, I still want my automation to continue working. And that's really introduced a lot of opportunity for building browser automation for the future, which is for actual like automation of work, not for the existing world, which is Mostly testing. And that's where browser-based has come in. And our framework Stagehand, which really kind of is a superset of the existing frameworks, Puppeteer, Playwright, and Selenium, it's actually much more durable than those frameworks Because it does a lot more fuzziness. And the kind of the advantage here is that it has an LLM functionality, an AI functionality built into the framework. In the old world, you can say, click the login button. It's the fifth button on the page. The color is red. It says login on it. In the new world, you can tell AI, hey, I want to click the login button. You figure that out for me. And if the login button changes color, maybe changes position, maybe changes from login to sign in, you can still find the button on the page using large language models. Large language models are very useful at parsing code. So as we think about what's possible now, we can now build much more dynamic functionality with headless browsers because we have LLMs that can on the fly figure out what the appropriate Action is to take on the page. And that's really increased the size of this market. In the old world, if you wanted to automate 100 websites, you have to write 100 scripts. In the new world with AI, you can write one script that can control hundreds or thousands or millions of websites by using these low-level dynamic primitives where LLMs can look at a Page and take that next intelligent action. That meets a demand for more browsers. ([Time 0:06:34](https://share.snipd.com/snip/d8415486-b6c0-42a8-adf4-a5ecd5a0e641)) - **Automating Invoice Handling** - A company repeatedly logging into multiple portals to download and upload invoices automated this tedious task with BrowserBase. - This dramatically reduces thousands of hours of manual work every week for them. Transcript: Paul Klein I talked to a company where a lot of their job is having a human being log into a portal, download an invoice of one place, log into a separate portal, then upload the invoice there. They do that thousands of times a week. And they're looking for a way to automate that. And BrowserBase, plus our frameworks and other large language models, all come together to let regular people build automations. We do sell to developers, so we are helping people build software. We're not, you know, a automation company. In fact, there are several automation companies built on top of a browser base offering more point-and solutions. For us, we really want to help power this next generation of builders, and in 2025, everyone is a developer these days. ([Time 0:11:43](https://share.snipd.com/snip/e1fad1f5-9ddb-4cdd-8741-804f5879ccdc)) - **Prioritize Automation Observability** - Use BrowserBase's observability features like browser recordings and logging to monitor automation reliability. - Observability helps identify issues quickly and maintain control of an AI-driven browser. Transcript: Paul Klein You really want to have observability over what's happening. You want to know what's happening when your agent goes to web page. We bundle browser recordings and logging so that if your agent's going off the rails, you're going to be able to see exactly what happened, triage it, and understand what you have to Correct in your application. You need to know what your AI is going to do when it has control of the web browser. We make it really easy to see that. And then Stagehand, Stagehand is a really incredible thing because it's model agnostic. It's an open source framework. So if anything, anyone can use it, right? It's quite nice. You can integrate it quite easily to your software and you get to bring your own models. We are not tied to any model provider. And this is quite helpful for our customers because when they switch, when the newest model comes out, they want to switch over to it and try it out. Well, thankfully in Stagehand, we have-party support for any model. So you can bring your own model and not be tied to any sort of model provider or even some sort of cloud host provider. You can use StageCham on your own and hopefully use BrowserBase with it. And of course, that observability story does tie into StageCham. When you use StageCham with BrowserBase, we have a lot of first-party StageCham observability to see exactly the prompts that you issued, the responses, the latency, the token costs. So we really make it easy for everyone to build reliable production web ([Time 0:15:24](https://share.snipd.com/snip/c1d31ecd-9574-4c99-b712-f18bb41af328)) - **Customer Trust Beats Partnership Tactics** - BrowserBase focuses on community and customer obsession rather than lock-in partnerships. - Trust and support differentiate them from larger competitors attempting to replicate their product. Transcript: Paul Klein Might need to think about a way to excuse ourselves and probably politely say this maybe isn't the best thing for us to talk about. So all in all, I think that this is something that happens in the valley paul graham has a great essay saying don't talk to corp dev i think it really summarizes this thing super well and It's common and part of the reason i addressed yesterday was one i think it's important that we talk about parser come out people are going to ask well what's the thing browser bases browser Infrastructure and aws agent core's browser tool? And I think the thing, first of all, you'll get from us is we have this observability functionality. We have stagehand. We have director. You have the full suite of tools. And secondly, you can trust us. We're not going to schedule a partnership meeting and launch a competing product. We're going to build alongside you. And every customer that uses browser base is extremely important to us at this stage. ([Time 0:20:19](https://share.snipd.com/snip/1708c2dc-4935-45d6-8cb2-9a693fc81866)) - **Food Stamp Rebate Automation** - Benny, a BrowserBase customer, automates submitting food stamp rebate claims via browser agents. - This saves users time on a complex, often neglected task by automating rebate filings. Transcript: Paul Klein Do is they have a great consumer app where you can take a picture of your receipt from using your food stamp. And Benny will use a browser, Powered by Browser Race, to gently go submit that food stamp for your rebate that you're owed. Something that's rebate is very painful. It's many different steps. And people who are going through this, they have a lot of stuff on their plate. They probably don't want to go get their like, you know, 25 cent or a couple dollar rebate. But Benny makes that super easy. And it takes this pain point that's relevant to certain people, creates a software around it, and then uses BrowserBase to power the agentic browsing behind the scenes that's actually Helping people save time. We've estimated that we've saved millions ([Time 0:22:10](https://share.snipd.com/snip/9954a2ec-726f-4606-8a64-2a14fb118ee9)) - **AI Browsing Like Human Driving** - AI agents driving browsers is like self-driving cars learning to drive in human traffic. - Rebuilding the web for AI is unnecessary; AI can learn to browse like humans within existing web infrastructure. Transcript: Frédéric Lardinois Web? Yeah. Paul Klein I'll kind of use another analogy. I'm in San Francisco. I took a Waymo to work today. And you think about AI driving cars as kind of like AI driving a browser. Now, the ideal solution probably is that Waymo should have their own lane. They should be able to drive as fast as they want. But that would require rebuilding all of this infrastructure and probably retraining how society works around handling cars, right? Instead, Waymo has learned how to drive among people and it does a pretty good job, right? We don't have to rebuild the infrastructure of the internet to help AI use the internet. We actually can use, you know, we can teach AI to use the internet the same way we do. And I think about like, I don't think websites are going to go away. I don't think we're going to rebuild the web because it's going to take generational change for people to stop using the internet. If anything, AI is going to get so good at browsing the web that we're going to be the slow ones. And websites might be needed as just like a human accessibility feature because AI is doing most of the browsing for us. And if you're old school, you're using a browser the old fashioned way. ([Time 0:25:04](https://share.snipd.com/snip/bbb9b60c-7e47-453f-8512-6ba386047cd0)) - **AI Leverages Human Creativity** - AI will automate tedious tasks, allowing humans to focus on what they enjoy. - Agents supervised by humans provide leverage, augmenting creative and meaningful work rather than replacing it. Transcript: Alex Williams Your thoughts. Paul Klein Let me repeat for him that, like, in terms of, do I use LLMs to think about strategy for browser base or like, how does it change your thoughts about being a developer? Alex Williams How is it, your perspectives about the work that you do? How has it changed your thoughts about society? There's so many elements to that question, but I'm curious from someone who's developing the technology firsthand about that role that LMS play in your life and in your work. Paul Klein Yeah, I'd say that I take the stance that really I see people operating with much more leverage. It's pretty cool. I don't imagine that there's going to be this doom loop that people are afraid of. I'm excited to automate the boring things I don't want to do, the tedious browsing that I want to do. And then keep me doing the fun browsing. I kind of like to shop for sneakers. I don't want AI to shop for my sneakers. I want AI to file my Delaware franchise tax. I do not, that does not give me joy. Right. And I look at a lot of the things that AI is automating. And you always will have a choice. You always will have a choice with AI will want to automate this for you. And if you can do it better than any sort of model can, there's a lot of value in that. I don't use AI to write very many posts of mine at all, because I actually think I really care about my tone of writing. And when I think about, you know, will AI impact the future of how people work? I certainly think it will. But I believe it's going to impact the future of work in a way that will let people do more of the things they care about and do less of the things they don't care about. I think that's this beautiful future that can happen with AI. I don't like to drive. I really enjoy just taking a Waymo and being on my phone in the back. That's way safer for the rest of the world than being on your phone while driving. And I get to kind of spend my time doing the thing I want to do, which is writing pretty funny tweets on Twitter, right? So this future to me is a future of much more leverage, where humans are having the power to supervise many AI and agents are able to do as much as you can, but still going to require the creative Energy that, you know, a person like you and I can bring to really accomplish that, ([Time 0:26:54](https://share.snipd.com/snip/f5b53b7b-3961-4893-b288-446adde1a040)) - **Leverage Expert Browser Infrastructure** - Building and maintaining headless browser infrastructure at scale requires deep expertise and ongoing effort. - Consider using BrowserBase to avoid reinventing complex infrastructure and focus on product-market fit. Transcript: Paul Klein An llm can fully replace browser base in standing up and maintaining and operating a headless browser infrastructure stack at scale. Now, I think if you want an LLM to build you a POC, I think that's totally possible. And I'd encourage people to use that. Use an LLM to stand up your own version of browser base. I'll tell you how. Get Chromium. Get Chromium trying to running on a server somewhere. As soon as you do that, you're going to realize that it's probably not going to be Amazon Lambdas because they're very resource constrained. So you have to go use Kubernetes. Okay, I have a browser running in Kubernetes. How do you scale that Kubernetes cluster? How are you going to do network isolation? My cloud code has a lot of trouble just fixing the bug. I don't know if it's ready for stateful distributed systems at scale. I think at one point it would be better for that. But what's kind of interesting is a lot of our customers at BrowserBase are actually people who have built this on their own organically with humans, and they switch to BrowserBase Because maintaining it, improving it, and continuing to innovate on that stack, on this piece of the critical infrastructure, is actually very time-consuming. It's not their subject matter expertise to be browser experts. That's us. And maybe ([Time 0:29:43](https://share.snipd.com/snip/ecb2aebb-5ff3-4481-81ec-87353ea4837c)) - **Cloud and Local Browsers Complement** - Local agentic browsers and cloud-based browser automation serve complementary roles. - BrowserBase focuses on always-on cloud infrastructure for running background automation when user devices are offline. Transcript: Paul Klein Well, we certainly don't view them as competitors at all. We actually, one of our customers, Josh Miller from Browser Company, was in a reply yesterday on the Amazon tweet. We are very supportive of the local browsing, the local agentic browsing initiatives. And where we come in and where we hope to partner with a lot of these companies is, what happens if my local browser is running, doing a task for me, for a consumer, and I close my laptop, Or I go, my internet goes out, or I want something to happen while I'm sleeping. You need something running outside of your device, running on a cloud, on a server that's always running. And browser base is really building the best browser infrastructure for those server-side background agents that are going to go out and automate the web on your behalf. So we think that there's a great parallel here where you can have your customer, consumer, local browser experience happening, but then as soon as you need to happen regularly on a schedule Where you're not going to be guaranteed to be at your laptop or connected to the internet, your background browser, your background agent can go ahead and take care of that task for you. And I think there's a lot of opportunity to power that. But in the end, those consumer browsers are just one use case. We're helping people build software. And sometimes to our customer's user, they don't even know they're interacting with the browser behind the scenes. They are just clicking more powerful buttons. And I think the future of software isn't actually you're thinking about browsers at all. You're thinking about more powerful ([Time 0:31:23](https://share.snipd.com/snip/3291d813-588b-4cd5-9605-75f26cf516a6)) - **Accelerate Developer Onboarding** - Prioritize developer experience by making onboarding and first API calls fast and easy. - Developer tools should make experimentation seamless and accelerate time to value. Transcript: Paul Klein I think grade abstractions are extremely important. And I'll kind of, I have a story about development experience. I think it's kind of fun. When's the last time you bought a jacket? Yeah, either of you guys. Alex Williams Last year. Last year. Paul Klein When you went to buy that jacket, did you go to the jacket store and you just kind of look at the jacket on the rack and say, it's got sleeves, it's red, it's got pockets, it's a medium. All right, I'll buy it. No, you try on the jacket. You got to know how it feels. You have to put your arms in the sleeves and make sure it fits well. And you're walking around a little bit with it, right? Alex Williams Well, that's the way I did it before I got married. Paul Klein Yeah, now maybe not. That's funny. But the jacket-buying experience, the trying on the developer tool, is the same thing our customers do when they use Stitch Hand, our browser automation framework. We have a great experience where you can go npx, if you're using Node, npx create-browser And what that does is it provisions a simple quick start experience right in your IDE. Set up with cursor or cloud code, you can code gen if you want. But we want to try and accelerate the time to someone trying out this product, understanding the ergonomics and making sure it's for them. So developer experience really is the most important qualification or conversion step in our customer's journey. And I personally measure what's our time from sign up to first API request as a key indicator of success of our developer experience. We want to make that as fast as possible. You need to escape from the browser into the IDE. That's super important for our business. And I ([Time 0:36:04](https://share.snipd.com/snip/c475e50d-978d-43c4-bbef-780746592f9e)) - **Open Source Framework, Managed Service** - BrowserBase open-sources its Stagehand framework to reduce web automation friction and grow the ecosystem. - Its core value lies in scalable, secure browser infrastructure operated as a managed service, not just code. Transcript: Frédéric Lardinois Open source for you and where does the enterprise product start? Paul Klein I think I draw the line in the way that makes most sense to me, which is that Stagechain is open source because it's a framework. It's something that you can use and you can run it with your own local browsers or browser base. But the value that we've based our whole business around is the value of the duration of the browser or the network bandwidth it's consuming. Those are our two SKUs that we sell, our two pricing models, is browser minutes and network gigabytes. So with that in mind, if I want to increase the number of browser minutes that are run on BrowserBase or increase the number of network gigabytes transferred to BrowserBase, what can I do? Well, I can unlock new market. I can reduce the friction to automating the web, which is why we built Director and Station to address two friction points we've seen with our customers. And I don't think, you know, Stagehand on its own right needs to be something that I charge money for because in the end, if I'm enabling more developers to automate the web and we're building The best browser infrastructure out there, especially one that works best with Stagehand, that is our opportunity to create value from that kind of framework. So I don't think it makes sense to open source BrowserBase, the infrastructure company, frankly, because it's not that valuable in the infra itself. If you have the infra, if I open source BrowserBase, I gave it to somebody, the value we create is maintaining BrowserBase, scaling it, proactively improving it, navigating all the Threats that can happen. We're operating browsers on behalf of customers or AI. A browser can go to any website. We have to be very careful with what we do, and we have a lot of interesting virtualization technology we use to secure the browser. The team watching over this infrastructure day and night gives teams that are larger and smaller peace of mind that they can trust that Browseways has that browser infrastructure Completely handled. That's something that's much more valuable than any sort of code. It is the ownership and trust that people put in us when putting mission-critical functionality on top of our infrastructure. And we deliver that every single day by acting as a sophisticated infrastructure company that's going to operate our browsers reliably and performantly for ([Time 0:37:56](https://share.snipd.com/snip/340e3bd8-cfef-4d10-bd25-a136eec79901)) - **Complexities of Server Browsers** - Running browsers at scale in servers is challenging due to resource demands and edge cases like time zones and codecs. - BrowserBase customizes Chromium and infrastructure using Firecracker virtualization and Go to optimize performance and security. Transcript: Alex Williams You developed? I mean, I'm curious what programming language you're using for that foundational technology, why you're using TypeScript, for instance. Paul Klein Do you work at AWS? Alex Williams No, I do not. You're a sponsor, but I do not. Paul Klein Well, I'll share with the pod. You know, let's talk about our technology, right? So browser-based, we run Chromium as our core browser. We only support Chromium. We think it's the right browser for our main web. On top of that, you know, we use a virtualization technology called Firecracker. Icracker. Alex Williams Because Firecracker has a very interesting history with headless computing, headless browsers. Paul Klein Yeah, virtualization is very challenging, right? Because not only like the browser itself is a bit of a virtual machine, right? It does a lot of isolation tasks. And the thing about the browser and Firecracker, and we use Go as a programming language, we use Node.js, a few other things. But I'd say the most interesting problem is that the browser itself was not designed to run on a server. You know, headless has always kind of been a hack. It's always been like a side thing for the Chromium team. And there really hasn't been much innovation on improving the browser to perform better in a headless environment. Once again, you think about the browser that you and I use every day, it's probably the most resource intensive application on our devices. It certainly is the thing that I'm running into the most when I feel like my back is slow. I have too many Chrome tabs open, right? And doing a lot of interesting work to customize Chromium, customize our browsers, and then customize the infrastructure around it to get the most performance out of our browser infrastructure Has really been something that we spend a bunch of mental calories on. And that kind of, those optimizations have been like subtle, but also really meaningful. And frankly, when you're running a browser in a server environment, you just run into a lot of weird edge cases. Like you got to have support for emojis. You need to have support for certain codecs. You have to compile the browser differently. You need to make sure that when you're running it in the right way, that you're setting the time zones appropriately based on where the user is actually located. Because if your browser is running in US East 1, but your user is in the West Coast, you need to make sure the time zone represents where they're at. Or if you have a browser that's running in one place, but the customer is in Germany and you have to set the right locale, so the page renders in the right language. These are all these little gotchas you experience along the way when trying to run browsers at scale. And there's just so many little edge cases that I personally experienced having worked in this headless browser space across two, three companies now for the last five years. So it's a labor of love. And as you continue to kind of like, once again, if you build this on your own, please send me an email how it goes. Cause I built ([Time 0:40:11](https://share.snipd.com/snip/76113719-bc14-43f1-a371-0e6aca123348)) - **Control AI Agents with Observability** - Use comprehensive observability including session replay and logs to monitor AI browser agents’ actions precisely. - Control agents by designing deterministic, atomic steps rather than letting them run unrestricted loops. Transcript: Frédéric Lardinois Buying a jacket for Alex or something like that, how do you make sure that doesn't happen, that they stay inside of those guardrails? Paul Klein Yeah. So we really have to start with observability. It's on our customers to help determine the guardrails of what is appropriate for their agent to do or not. Because we're a platform, I can't necessarily say that for all customers, but when we think about this, we're often like identifying, hey, is a browser going to a known malware site And flagging that for our teams internally? We want to make sure that we have good, you know, safe practices for our browsers overall. They're all of course, completely sandbox, we don't share any hardware across browsers nor do we ever use any browsers they're all destroyed after being used but for our customers The way that we can help them actually steer their agents appropriately is by giving them best-in observability a full session replay where they can watch exactly what happened log Lines around what actions were taken in the browser you know network requests they can see what network requests were made from each page. Being able to analyze that is extremely helpful for ensuring your agent is doing exactly what you want it to do, which is important in this age of non-deterministic agentic software. This a little more deterministic and that means isolating the llm calls into single atomic steps instead of telling your agent go buy me a jacket and letting it run loose you tell it go To this big vendor find a jacket like this and buy it with these credentials assuming that it's no less than this price and you break it up when you have one prompt that controls a browser And just kind of is in a while loop you're going to have non-determinism when you have several atomic steps that are verifiable, cacheable, and explicit, you have a much more reliable Browsing experience. Frédéric Lardinois That's interesting. One thing I was thinking too, and maybe to wrap it up, but as I'm ([Time 0:43:08](https://share.snipd.com/snip/b8fae4d4-efa4-44de-abad-1d80446b1e29)) - **AI Navigates Complex Web Pages** - AI agents can navigate pop-ups, ads, and consent dialogs effectively, reducing manual browsing hassles. - Combining vision models with text-based models improves automation of image-heavy or complex webpages. Transcript: Frédéric Lardinois How do you handle that? Do you run ad block on the headless browser? Paul Klein There's an option to turn on ad blocking if you want to, right? Or cookie can send pop-ups. But generally, I think that agents are quite good at navigating ad blocks or pop-ups or any of that stuff. It's actually quite easy. Where they have troubles, actually, when there's a lot of information and things that aren't necessarily in the DOM and HTML. If you have a very image-heavy website, using a vision language model, a VLM, it's very important because you're going to want to be able to see what's in that image. If someone is presenting, I've seen this happen on some of our customers' sites, they go to a place to buy something and all the product information is in an image. Very hard for an LLM to parse an image, right? So you have to use a combination of vision plus text-based models to really be able to automate the web reliably. And I think one thing we aspire to do with stagehand is that you won't have to think about that when you use our browser framework. We'll let you provide multiple models, and we'll pick the best model that works well for the page that you're automating. A developer shouldn't have to think of all these tricks. They should have a browser tool they integrate like Browser Race and Stagehand that just works well for their agent and they can focus more on their agent reasoning logic, which is really The proprietary IP that's important to their business. Awesome. ([Time 0:45:08](https://share.snipd.com/snip/a6133886-f1f6-4a4f-ad0e-3d064e8a2456))