Allie Howe Hey everyone. Welcome to today's episode of the Insecure Agents Podcast. I'm your host, Allie Howe, and today we have Peyton Casper with us. He is in charge of identity and trust within the product department at Browserbase. Peyton, would you like to introduce yourself?

Peyton Casper Yeah, for sure. Number one, thanks for inviting me. I'm really excited to to talk with you guys today. You know, identity and trust I think is a it's a big concept, especially like within kind of the growing AI space, right. And how do agents represent themselves on the web and then also within the broader ecosystem. And so from that perspective, you know, that's kind of my focus at Browserbase is how do we make sure that when our agents or customers agents are out there on the web, you know, they're not abusing downstream websites. They're, you know, giving off ideally like an identity that, you know, we can all trust and then also partnering with various standards, right, that are, are currently going through various bodies.

So that way we all speak the same language when these agents are out there operating and and we can recognize them effectively.

Allie Howe Yeah, I feel like I've heard more and more lately in the last like 3 months at different conferences and events and stuff, this concept of trustworthy AI really starting to become mainstream. I applaud Browserbase for even having, you know, a title of a role for identity and trust and regarding agents. And I'm, I'm sure they've had that for, you know, not just three months, but more than that. So super exciting that they're focused on that and that's what you focus on.

Peyton Casper Exactly right. I think like a lot of the AI companies are still new and, and when that happens, obviously the focus is on developer experience, right? And I think sometimes security can be antithetical to that at times, right? It's striking that right balance of how do we make sure that we can still move quickly and we can still, you know, develop cool new features and stuff like that, but also, you know, be good stewards of both the enterprises that we have to work with, right? Because they have their own security implications, but then also of the broader Internet as well. It was an interesting hire from Paul, first product person that he brought on within the company.

So I think that says big things about where his mind is at from a company perspective as well.

Allie Howe Whenever I think of Browser agents, I think of you just because I followed you on X for so long. I know you're working with Browser Agents in your own startup, Coffee Black. Prior to joining Browserbase, what learnings did you have from Coffee Black that you brought to Browserbase?

Peyton Casper The biggest thing, one of my largest customers at the time was a property management company and they wanted to build these agents that go out and collect utility bills from various utility providers that might be PG&E and here in California, Ameren in Missouri, etcetera. And so the biggest challenge we ran into wasn't even the fact that, you know, a cool model couldn't click the right button or couldn't fill in the right field, which is what you see a lot of the hype around it was how do I get credentials in securely and then how do I get past anti-bot, right? And so we're in this weird period in time where historically bought traffic has been looked at as fraudulent, as bad for websites.

And there's a long stream of reasons for why that is right when you look at the external marketplaces that get built, buying and flipping various items that are in hot demand like Ticketmaster tickets and things like that, as well as like warming up accounts and, and scraping data, right? But now we have these agents that are are starting to become productive, right? As kind of the capabilities of these models continue to increase, we're seeing more enterprises say, what can we do to actually replace kind of repetitive manual labor that's happening already. And so that particular task was literally they would get these bills in the mail and they would have somebody type in all that data into a system, right?

Which is really like archaic process. And so really great use of browser use agents. And I realized like, hey, the biggest challenge here was actually nothing to do with models. I think it was a pretty clear sign that like it's time for the industry kind of figure out what is our path forward there, right? So that's how what led to the conversation number one, me joining Browserbase and then to, you know, our initial partnerships that we've announced with Cloudflare, with Stitch and a number of others essentially.

Allie Howe Yeah, I feel like those partnerships are key because it takes everyone to implement new standards and to adopt new standards across the industry. I think, you know, we're seeing that with like MCP, for example, but there's new protocols as well. So just like web bot auth, I know, which is new. I think I want to go back to what you said around browser agents are sort of following the same trap that bots were where they're well meaning, but they're getting blocked by different websites. So I need some sort of a new solution to allow these agents that are well meaning to state who they are, what they're trying to do to gain access to different websites.

How do you see web bot auth being the solution for that? And why do we need web bot auth in the first place?

Peyton Casper Yeah, historically, if you look at the way that bots or headless browsers kind of represent themselves, there's something called a user agent in the browser. And this is just a string, so, and anyone can set it, anyone can change it, and it's basically set at a browser level, right? So if you were to look at it on Safari, it would say, you know, roughly Safari, Gecko and a version number and a few other things, right. On Chrome, it would say Chrome and Webkit and a few other stuff. And that was OK in the sense that, like, it was a way to identify sort of, you know, hey, this is a browser or this might be a headless browser, right?

As that evolved, techniques on both sides advanced. So people got better at masking the browser if it was trying to act like a headless browser, as in there wasn't a human actually running it. And on the other side, obviously, the companies that were protecting these websites got more advanced. And so famously, this led to Captchas. And we all know them because you have to click the right squares and things like that. But now they're actually evolving to where they're invisible. So the challenges that you see actually don't really do much. It's really more of an artifact of kind of the old. And now they're really just monitoring various data points about the the browser that they see.

And I think a really cool one that that we talk about a lot internally is these anti-bot companies will actually render a canvas invisibly in the browser. And they will benchmark how long it takes your GPU to do that. And they'll use those metrics to basically identify do you look like a Mac or not, or do you look like a Windows, right? And so the reason why that's so effective is these headless browsers don't have GPU's. So when they go and try to render it #1 it may not even look like come out the same, but two, it takes much, much longer. That falls out of that profile window, right?

And so we have these varying kind of cat and mouse games where someone comes up with a solution, someone comes up with another way to detect it. It's just it's really, really inefficient at the end of the day, right, It's a lot of wasted engineering resources. And so our perspective is really simple that the market for reliable enterprise automation or browser use agents or whatever you want to call it is only going to continue to grow. These models have not shown that much of a decline in their ability to continue growing relative to their capability. And so the size of that market is just going to continue to grow.

And if we come to the table and we say, hey, we're willing to identify ourselves with this standard called Web Bot Auth. And the way that it works is you basically sign a header payload with a public private key pair. And we would advertise or our customers would advertise or another company would advertise a public key at a, you know, well known endpoint. And then we would sign the header payload with the private key on the back end and include that on a request. And so when reCAPTCHA gets that request, they can basically take that sign header payload, They'll see the URL, they'll go to our public key directory and validate that it was signed by Browserbase or it was signed by our customers.

And so this partnership, you know, I think when Cloudflare and Browserbase announced it, there was a lot of like buzz around like, oh, Cloudflare is like trying to shut off the Internet, right? Or, you know, kind of gatekeep the Internet and you know, I think it's it's missing a critical point where number one, we need this like verifiable way to understand like who can we trust, right? Because otherwise we're constantly in this cat and mouse game. But then the the second aspect of it is that anybody can implement this protocol, right? You can go to the web bot auth GitHub and you can expose the public key directory and you can go talk to Cloudflare, go talk to DataDome or reCAPTCHA and say, hey, this is our use case.

We think it's a legitimate use case. Let us in, right? And I think the anti-bot companies are taking a measured approach where they're saying, hey, we're going to kind of section off this traffic and we're going to trust but verify. So we're going to let it in, but we're going to monitor you right to still ensure that even though you're signing this traffic, you still shouldn't be able to D dos the site or make 1000 purchases of the same item, right. And I think most reasonable humans would agree with those those limitations, right, because they disrupt kind of the implicit contract that's there with the site owner.

Allie Howe I. Was wondering that as well. If Browserbased signs you know this agent and this agent can authenticate like and say who it is. Doesn't mean that the person controlling that agent is, you know, doing right things at that age and maybe they're up to something malicious or they are trying to DDoS aside. So it's nice to see that there's mitigations in place for that as well.

Peyton Casper Yeah, can be interesting, right. I, I think everybody's a little bit nervous. Everybody's kind of like we see like the need for this coming. I think it's going to, obviously it's continuing to take time to get everybody aligned and, and just kind of there's a off the books proposal going through IETF right now around how can we have like an abuse monitoring endpoint. So effectively, if we run into your, your signed header payload and we, we detect that it's bad, right? We want to give the site owner the ability to report abuse. And so that would like give like real time feedback that like, hey, you know, this bot isn't adhering to standards, right?

So I think there's a lot of really cool things and and I would obviously encourage like anybody listening, all of these discussions are pretty much happening on the forums for this proposal for IETF. Obviously, you know, I think the technology moves a little bit quicker outside, right, as it should and while the standard is being developed, but it's been very positive so far.

Allie Howe I feel like so much infra today uses shared IP addresses as well. In the past, you could, you know, effectively say, OK, this IP address had malicious behavior, let's just block this entire IP address. But then you might be blocking, you know, hundreds of other agents that are that are good agents. So I think that also highlights the need for web bot auth today as a new solution.

Peyton Casper Yeah. I'm curious, Allie, have you looked at like AP2 or any of the payment protocols?

Allie Howe Yeah. No, I haven't.

Peyton Casper OK. Yeah, I think like it's, it's interesting like a lot of companies are like starting to throw out like those different like protocols. But interesting like aspect that comes out of that is something called a mandate model, which basically tries to ascribe like a permission model based off natural language. So if you say, hey, this agent should only be able to buy these tickets, right? Or should only be able to navigate to this website or shouldn't be able to delete this organization within a website or an application. I think that's like a really interesting additional problem, which is OK, once we know like what this agent is, how can we like actually prevent it from doing bad things, right?

I know this is like I've seen on X, like you've talked a lot about how do we prevent like prompt injection? How do we prevent these agents from coming off off the off the deep end, so to speak. So yeah, I'm curious your thoughts like on how do we like apply like permissions models at scale for agents?

Allie Howe I was just at Zenity's AI Agent conference in San Francisco a couple weeks ago and Michael Bargury, their CTO, gave an amazing talk about the need for hard boundaries instead of soft boundaries. I think over the last probably like 6 months, we've heard a lot of talk and enthusiasm over LLM guardrails, which I still feel like are a good mitigation to have in place. But I think it's just one defense in depth solution, part of a much larger picture. And guardrails, he said, are soft boundaries. They're not hard boundaries. There will be a prompt injection jailbreak, something that eventually will get past those guardrails. They might block most of it, but all an attacker needs is to be right just one time.

And so in order to feel more confident in our security posture, we should actually be more focused on hard boundaries. So I love what you said about creating from the permissions model around, OK, this agent is not allowed to buy these tickets specifically or this agent is not allowed to delete this resource. And having a hard boundary there, like permissions based boundary within like the infrastructure or the networking or something where like if the agent was manipulated and someone, which is so easy to do. And that's what my talk was at this conference was about like goal manipulation at this. It's so easy to get these agents to do the wrong thing.

So we just need to like plan for that and basically assume breach. And so when they do misbehave, like there needs to be a hard boundary, hard stop in place. There's lots of different companies and I'm really excited about that are developing solutions there. I think Keycard Labs just announced this week their round of funding. They're trying to create context aware auth, which I think will really help with the hard boundaries problem.

Peyton Casper That's interesting. I actually saw like obviously with the Atlas announcement, someone put up a website where they on the website they told the agent to do something different than it was told to do. And they got like Atlas to basically like re navigate itself away from the website effectively. Which I think, you know, is a whole other kind of worms of like if these AI like agents or browser use agents are susceptible to the instructions that are on a website itself, You know, they can run into anything on the web, right? Which is also very much very human like in a way when you, you know, look at like kids and stuff like that.

Very susceptible to suggestion. But yeah, it's a very interesting place for sure.

Allie Howe So out of curiosity, how does Browserbased view things like Atlas and those AI browsers? Is that like a competitor or is that like a totally different thing? Like why would someone you use want to use like Browserbased versus Atlas? Like is Browserbased more secure?

Peyton Casper I mean #1 I'd like to say we're more secure. We spend a lot of time thinking about like access control, credential management, and obviously identity. But I think that the biggest thing is, you know, an AI browser is meant to run on your laptop, right? And so it's supposed to be a convenient app, which is great for things that you want within your, what I would consider context window, to steal a phrase, where there's a limited number of things we have in our mind at any given point, Browserbase is really, really focused on, OK, once you have like 5 things that you need to happen maybe 100 times or 1000 times, right?

Some of our use cases are compliance verification. So going out and like researching a particular person or an enterprise, things like that. As well as like receipt management for like expense tracking apps effectively, right? When you make a purchase, you need to upload that receipt. We can go out and fetch that for you automatically. There's a whole host of like other things that those use cases where they need to happen repeatedly. And that's not something that runs on your laptop, right? Same way that we have kind of local software and we have software in the cloud. It's just a matter of scale. So I think that's our first criteria.

And then what we bolt on top of that is of course that security, which is OK, now that you don't own the hardware, how do we make sure that you can trust this, right? How can we make sure your credentials as much as possible only live on your device, They don't get stored anywhere else that this agent isn't going to, you know, get prompt injected to spend $1,000,000, You know, that it's not going to be abusive to, to website owners and things like that. So I think that's kind of the the core difference between kind of the AI browser wave, which is I think aiming to be that next generation tool set for users and how do I go from this like automated workflow and scale it up 100 times or 1000 times?

Allie Howe Like how you touched on the hardware piece as well, because I've heard a lot of talk around different needs for infrastructure for agents, how they need to be in like private sandboxes or there's companies that are building like basically their own, like clouds specifically for agents. Are you seeing like different infrastructure challenges with your work at Browserbase?

Peyton Casper Yeah. I mean, I think the biggest thing is where do you run like Agent code, right? So we're actually, we just announced with our Director release something called deployments or functions. Basically what that means is when you come to Browserbase, the code that directs the browser that tells it to click on this button or to navigate to this page or fill in this field that runs next to the the browser itself now. And that was like a huge lift because if you've built on top of Browserbase before, pretty much had to maintain the kind of code runtime that I actually like talks to the browser.

And that's a whole additional stack of hardware that you have to maintain. And so like you obviously see startups like E2B, Vercel functions, Lambda functions, I can go on. So I think there's that side of it, which is how do we have these like lightweight execution environments where like a model can like iterate on top of. But I think the, the other like big shift is historically all of our software has been built to be synchronous, right? Like we've optimized for low latency, we've optimized UX for human driven access. And now that's changing, right? And you can see this in like how users are using tools like codecs or background agents with cursor or, you know, insert your, your option where you might fire off two or three tasks and then you might switch over to another rebuild and fire off another task.

And all of those are happening in parallel. So the user's interest actually no longer like, you shouldn't really care how fast something happens because you can do many things at once, right? And I think that that perspective shift that like design shift is a really big thing, really hard thing to design for. Because now it's not like, oh, I sent a request. It's like, well, I kicked off a request at some point in the future. It could be in 5 minutes, it could be in an hour or could be in a month. I want to come back and get the results of it right? And then what if the agent runs into a problem right?

I think for browser use agents in particular, what happens if they run into a 2 factor auth code? What if they need credentials right? How do we bring the human back into that loop where they in a productive way, in a non annoying way, can be part of that process? I think we're still learning. I don't think anybody has really like nailed that, but that user experience and then the corresponding like software changes of like, how do we queue events properly? How do we make things, you know, fault tolerant? How do we request permissions, you know, out of band effectively? Those are really, really hard problems and and we haven't had to worry about them.

And I think the final piece I'll add, and I'm curious your thoughts too, is like when you look at like the evolution of CPUs, we had multi core architectures for, I would argue like probably about eight years before the vast majority of the software industry really even started to leverage them. And even then it was the gaming industry that sort of led that wave. And now like you Fast forward today and we have consumer CPUs that have 8 cores. So now it's a little bit more well known, you know, AMD kind of famously LED that charge. It's going to take time for the industry to mature around that user experience.

I'm curious like have you seen specific like hardware changes or specific companies tackling this?

Allie Howe Yes, I'm familiar with Daytona that's in this space. I think they've got an interesting solution, it sounds like and they've run into some interesting use cases where agents need their own sandbox to run both for just like standard, like infrastructure reasons like CPUs, workflows, all that, but also just flying from a security standpoint. If you're going to be doing like a high risk function and like processing certain like data and you want to be like in a sandbox, you're going to do like code execution. That maybe was influenced by some sort of prompt the user gave you or code they actually like sent all the way through.

Like you'd make sure for sure, like any high risk functions like done within a sandbox. I feel like that is definitely a like strong use case for like new infra or agents. But yeah, I think I've listened to one of their podcasts they did recently on the the Infrapod and they were saying that there's like a need for an entirely new cloud just specifically built for agents. I don't remember specifically like all the reasons why, but I remember walking away from the podcast thinking, OK, well, I wasn't thinking about that. I didn't realize there was that strong of a need. They were that different. But I, I, this is a recurring theme I'm just seeing with agents, OK, we're rewriting how we're doing auth or rewriting how we're doing infrastructure.

I don't think we're totally rewriting how we're doing application security, but I mean there's definitely new things there as well.

Peyton Casper It's weird, right? Because like all of the, I mean, on the security side, just focusing slowly on that, like all of our identity management has been focused on humans. And now it's like, Oh, well, we kind of have these like pseudo humans, but they're not like intelligent enough where you can actually give them like their own like autonomy really. And then that's a shifting target, right? Like OpenAI could come out tomorrow and like completely change the game on what they're capable of. Now all of a sudden you want them to have access to your bank account. So I think, yeah, it's a it's a interesting space to be in for sure.

Allie Howe Curious to hear your take on this, but our capaches like, you know, Cloudflare turnstile, where you go click the box and it says, you know, click this box if you're human. Is that dead now just because agents can do that?

Peyton Casper Yeah, I mean, for the most part, right, Like the clicking, like interacting with the web pages, it was solved I think with vision models candidly. But even before then, like most, pretty much all the providers just used invisible. So they kind of kept the capture part around because I think it makes humans feel like they're part of the loop. But all of the actual like validation is done solely off of what does your browser look like? What does the machine look like that we can, you know, see what IP are you coming from reCAPTCHA particular which really, really excited to announce that like we're going to they're embracing the Web Bot Auth standard alongside Cloudflare, Stitch and ourselves.

You know, they're focused heavily on on identity, right? And the way they do that is, you know, they're Google, they have Gmail accounts and, and things like that. So they have a pretty good idea of this browser or this device is being used by a human versus this is being used by a bot. And I think those providers in particular are going to have an interesting time trying to adapt to this agent world where it's like, OK, not only do I have to kind of let this bot in, but I also need to pass through the identity of the user that I'm operating on behalf of.

I've had some conversations on the Amazon side where they want signals about their consumers that are purchasing, right. So they have, you know, it's a growing, it's it's small, but it's growing agentic commerce traffic that they're seeing right now. The agents just kind of look like black boxes. And so if you're a merchant, all you're seeing is, hey, agentic commerce agent number one has come in and purchases item. You have no idea who's purchasing it or anything about your target audience, which I think is good from a privacy perspective. But I think there's also that balance of like if you were running a store, you kind of want to know generally like, who am I appealing to, right?

Like who is buying my stuff a little bit, maybe not like names and addresses, but general categories. And so I think we basically inserted these agents as like proxies, right, for for a lot of things. And now everything that we're doing is trying to like say, OK, well, all this historical stuff that we have, we kind of need to incorporate that concept of a proxy now.

Allie Howe Yeah, it's interesting. And I wonder too, what happens if you're a well meaning user and you created a browser use agent and browser use agent like goes off the rails or falls back into a prompt injection or does something that you did not intend it to do, Does that fall back onto you like your identity is associated with that agent? Like how do we determine, I guess like this was a mistake versus this was malicious on purpose?

Peyton Casper It's a great question. Actually. I think that's a, it's a really hard question. End of the day, I think it's shared responsibility, right? So obviously as an infra provider, we have a responsibility to make sure that like our platform is secure, that no one can come along and hack the agent runtime or anything like that. You know, if you're providing an agent on behalf for users to use, you have a responsibility that like as much as possible that agent can't be prompt injected, right? I think when you talk to enterprises, obviously they want to deploy workflows that automate tasks, right? Like I don't think anybody doesn't want that.

But I think their chief concern is like exactly that. Like the humans I trust to do this job, they have a lot of responsibility, right? They have a lot of power. And if we extend that power to this autonomous being, how do we make sure that that doesn't get abused effectively? So I think companies as they have been historically own some of that liability as well the, the agent developers and I think the users or the downstream organizations do, right. Historically, we've kind of just given tools like Terraform, CI/CD, kind of these admin permissions where they can just do whatever they want. And that's probably not going to fly in a world where you also give this agent a credit card, right?

Or you, you know, give it access to 8 different accounts. How do we define that permissions model in as much of A deterministic way as possible? I think it's really important. I haven't seen anything that I'm like, yes, this is like the answer just yet. But I do know like a lot of really smart people much smarter than me are like working on that problem because I've talked to them and they have some like really crazy ideas. I'm excited to see where that pans out. But I think there's going to be that balance of like deterministic versus non deterministic, right? The benefits of intelligence and that it can adapt.

But then, you know, how do we put those those guard rails, to your point, how do we put those hard stops in place in a truly deterministic way that's also like extendable. So when we're talking about the Internet like APIs and MCPs, I think they have some benefit in the sense that they kind of can bolt on to the existing authorization layer that's in the API when you're dealing with a website either have to reinvent RBAC or you have to tie it into the credentials. So you say like, hey, we're going to only give re access to this particular agent, right? But I think there's a long tail of like really old websites where there is only the credential and that credential can do anything on the account, right?

But I think there's progress being made and I think the constitutional verifier work that like Anthropic has been doing is probably going to play a big part in that into how can we like guard the agent from, from injection. But I think it's shared liability. And I, I think that's the case for everything. Waymos are famous for this where it's like if Waymo gets in an accident, whose fault is it right? Well, you know, in the beginning they were really, really cautious and they have telemetry on why did this self driving car take this action, right? They have recordings of where was it out in the street when it got hit?

What was the speed of the the Waymo? What was the model decision? What was the input data? I think agent telemetry is going to be just as important and it's also candidly a big focus of what I do, a Browserbase as well, which is how do we do that in a Safeway, but ensure that, you know, our customers don't go out and abuse other sites. And part of that is giving our customers the tools to understand what their agents are doing, and in some cases, you know what they shouldn't be doing.

Allie Howe Yes, for sure. Yeah, telemetry, observability, huge for monitoring your AI application, making sure it's doing the right things. Yeah, I feel like we kind of touched on it because I wanted to ask, like if I was a developer and I wanted to go make a browser use agent for the first time using Browserbase, what are some things I should be thinking about in terms of like both performance and security? And I think we talked about, you know, permissions model, making sure that permissions are scoped appropriately, making sure that it's aligned and it's only supposed to, you know, accomplish a very specific task, not just do a broad range of things be connected to like 5 different bank accounts or something, for example.

Anything else that a developer making a browser use agent for the first time should be thinking about?

Peyton Casper Yeah, I think like the the biggest thing, you know, we have a an open source project called Stagehand and and candidly, I think it's probably one of the best ways to build these browsers agents because you write a natural language prompt, you say click on the sign in button. But then there's a caching mechanism built in where once it executes correctly the the first time, a cache is kind of the selector that it interacted with, right. And so the benefit there is that rather than building like this super prompt that tries to like adapt to like any environment, you're actually building a, a set of steps that need to happen.

And when we talk about like repeatable, scalable automation, that is really, really important to making that work, you know, 1000 times over. So I think, you know, looking at looking at them as like deterministic, like repeatable flows is important. And then, like I would encourage people to play around with the standards like Web Bot Auth and others, I'll speak for myself, like growing up, right. So all these things come across kind of on Hacker News or you saw them come across on Twitter, right? And you can feel like this, like, you know, Hey, there's these people out there that are like doing this thing and it's going to like impact the, the broader Internet, the end of the day, like they're just people, right?

Like I, I sit in an office in San Francisco. No, like we're all just people. Like you have an idea or you have ideas of like how we can solve some of these problems, like get involved. Well, that's happening in the open. And I think like that's like one of the benefits, one of the perks of the tech industry. So much of it does kind of happen in that open space.

Allie Howe I agree, yes, I've noticed that as well. There's so much room for collaboration and for people to get involved. And like you said, we're all figuring out together. This is our very first time, as you know, I guess the society like implementing AI agents. And so yeah, no one's an expert here necessarily. Like jump in, let's do it together. I love that spirit. Is there anything else you'd like to add about what's next for Browserbase, what you all are doing, or anything else?

Peyton Casper I would just say, if there's anybody listening, anybody that's focused on how do we kind of help separate good bots and bad bots, right? How do we create that fast lane for, for bots that are, are valid use cases and trustworthy. And then if you're focused on credential management, right, if you're 1Password or something similar, right, that we've announced a recent partnership with them, or if you have like crazy ideas about like RBAC and, and how we apply that to agents, those are focuses for me. I would love to talk to you from a Browserbased perspective. Obviously, we have a lot of things coming down the pipe in terms of how do we make browsers faster?

How do we make them lower latency? How do we continue to push forward like the standards that, that we're pushing our Director products is a great like manifestation of kind of like what's possible, right? You know, it allows us to take all of the the browser infra that we build and I actually like build a product that our customers build for their customers, right. So that dog fooding that we use to make our underlying infrastructure product better is is really important. Last like thing that I would say is like we're hiring, I think pretty much everywhere. So if browser use agents sound cool, if kind of working on identity sounds cool.

If talking about anti bots is your thing, definitely reach out and we'd love to chat.

Allie Howe Amazing. Where can people reach out if they want to join in on those conversations?

Peyton Casper Reach out on Twitter, you can send me an e-mail. It's literally my name at browserbase. com. Please don't spam me. And then you know all the standards, conversations and Stagehand. It's all open source. If you just search Web Bot Auth, you search Stagehand. It should be up in the 1st results that come back.

Allie Howe Well, thank you so much, Peyton, for coming on today and bringing your expertise. I know you have so much background and browser use agents. So thank you so much for everything that you've taught us in our audience today. I'm super excited to stay in touch with you and Browserbase around how you're solving identity challenges for browser agents.

Peyton Casper Thank you, Allie, and I appreciate the invite.