رونوشت کامل Building the Cloud for AI Agents with Cloudflare's Dane Knecht and Rita Kozlov

to... I mean, one thing that we do very much have in our DNA is innovation, having retaining an ability to have agility and flexibility when things come up. And so I, I think it was a combination of we have the right foundation, but also once, you know, new, uh, protocols like MCP came out or we started paying attention to harnesses, like, we're able to adapt very quickly to whatever is happening in that moment. Dane: I mean, if you think about it, the whole cloud transition that everybody has been told to do for the last 20 years, uh, it really was just shifting from a data center to another data center you know, data center you own to a data center owned by one of a couple people. Um- Yeah, Corey: that's the lift portion of the lift and shift. Yeah, yeah. The shift, eh, we'll get to that later, as soon as we finish migrating off the mainframe. Dane: I think @S3 was a, was, was a new kind of primitive, uh, uh, but then it kind of regressed from there. Then it went back to just like, "Oh yeah, let's just give you a full machine and, uh," basically everything that you had, uh, in a data center before. Uh, never kind of going back to, uh, you know, the core principles of like what are you trying to accomplish? You're trying to accomplish efficiency. You know, how can you make sure that every machine that you have is running, you know, as close to capacity at, at a given point? Corey: It- it's weird in that things are speeding up incredibly, but also at the same time, it feels like there's a, there's a drag effect on it too. I remember a few weeks ago during your birthday celebration, uh, which is an interesting thing for a company to say, you released this CF command line tool, which, uh, solved between Wrangler and your other one. I forget what it was. They were constantly at odds with each other. Hey, a unified experience. And I remember giving that to my agent the first day it came out. Like, "Hey, use this instead. Cloudflare just released it." And like watching its thinking traces where, "I don't know, sounds like horse shit to me, but I'll give it a shot." Uh, and it also was trying to deploy things on top of Cloudflare pages, which is largely been deprecated and not the way that you folks envision these things going on. But getting the new way of doing things into the, into the weights and getting the old way, the old ways of doing them out when there's 10 years of historical blog posts on it being baked in feels like it's an incredibly difficult challenge. Dane: There were some arguments there. Uh, I, I was like, "No, it needs to be called Wrangler 'cause that's what the other ones know. Uh, uh, it needs to just be Wrangler 3." Uh, and, and they're like, "No, no," like, you know, s- "We should have called it CF all along. That's better. Uh, and we're gonna do it in a way that, you know, the agents will still just prefer it. It'll just be such a better experience for the agents." And I, I think, you know, been, it, they've been proven right. Like, uh, you know, first of all, CF's a lot easier to type, but, uh, the agents seem to naturally be able to understand it and, uh, it was kind of built actually more for an agent to use than a human to use. Corey: Oh, the, the search operator if I... Where you explain it in plain English the thing- Yeah ... you're trying to do and it spits it out? Awesome. I love it. I'm still trying to figure out how I can exploit that. You know, write Python code for me because why not? Everything's a database if you hold it wrong. The other side of it though, I, I have to say Wrangler was a bad name in that you go six months without touching it and, "What is this Wrangler thing installed on my system?" CF at least you can trace down and sort of start to understand a bit more effectively. Legible to humans and legible to agents does seem to be converging. Rita: We hired an engineer who had just moved to Austin, and they were like, "Yee-haw, Wrangler." Um, and, you know, you You find these things endearing and then like seven, eight, nine years later, you're like, "So-called Wrangler." Corey: It, it's weird. The, the casual decisions we make, like, "Eh, I'll go back and fix that later," and you realize inadvertently it's like, "Yeah, I'm just gonna replace the entire foundation in that building," and oh no, suddenly, well, I guess that's what we're calling it forevermore. Rita: But it's interesting y- there's, uh, there's both sides to it. On one hand, there's everything that LLMs know, and obviously it's hard to displace that. On the other hand, it's so much easier in some ways to get LLMs to adopt something new if you can be a bit intentional about it because you can just give it a prompt and it will follow that prompt to the T, whereas humans are like very bad at following instructions, and we do fall back to old habits much more easily. So I think that that was something that the team put a lot of thought and effort into that helped make the transition to CF so much easier, like the fact that you can give it a single prompt. And, um, before even migrating between versions of Wrangler, that was something that I I would very painstakingly come to a customer and be like, "You guys really should do this." And they would be like, "I don't wanna fiddle around with this for a month. I wanna do other stuff that's actually important to my roadmap." And now it's like, "Just give it this prompt." And they're like, "Yep, I'm done." And we did this with a bunch of teams internally. Corey: It, it's a near certainty that there was an internal debate around this, so I wanna make sure I call out something that I like. Um, every provider now has skills you can give your agent that winds up explaining how to interface with the various things that we're using, how to, how do you contextualize this platform that you've suddenly gotten exposure to. And a, without naming names, a, a lot of them are basically SEO for agents of, "Oh, any time

someone wants to deploy it, make sure you use this." So I went through yours, and it is remarkably even-handed across the board. It is, "If you are doing this on Cloudflare, then go ahead and do this other thing." It didn't turn into, "And remember, Cloudflare is the best," which some other skill packs do. And that has been wildly annoying to me when you, especially, God forbid, you have two of those things installed. So as fun as it is to wind up yelling at companies when they get things wrong, I do like to take a point every now and then of m- making it obvious when you've done something right. I'm sure someone was like, "Oh, think of the growth hacking we could achieve if we just basically sell our customers for glue." Dane: Well, I mean, I, I, I think that's pre-baked into our DNA. Uh, you know, things like even just from privacy to, uh, transparency to, uh, the customers are not our, our, our data and, and, you know, so, uh, that is not like a tops-down thing that, you know, that is kind of baked into, uh, uh, the people that we hire. Uh, there, there was no like conscious decision. I'm glad to hear that. I'm actually excited to hear that you, you say this, but, uh, uh, I actually didn't even know that. Corey: I was expecting that some providers would absolutely let marketing run loose inside of those skills, and- Rita: And some do ... Corey: less than I was worried about originally, though, which is kind of awesome. Like, one other thing I'm realizing now is that we've been begging as humans for the internet to be better in various ways for 20 years, and only now that agents are a thing are companies actually getting off their corporate asses and doing something about it, where, oh yeah, turns out that we, if we stuff like eight megabytes of JavaScript into a two-kilobyte instruction file, agents are just not gonna click on the ads or read it anyway, so we should just put the text on a text file and make that available. It's like the second greatest thing to happen to the internet. Like, it's, it's reminiscent of GDPR day, when a bunch of companies were suddenly unprepared, didn't know how enforcement was gonna work, and for a few days they turned off the JavaScript and the entire internet was blazingly fast. It's like, ha, you coulda had this, but telemetry is important. Rita: So you know what's funny about that is on GDPR day was the first time I was- Initially the original Workers PM and it was the first time that we saw a jump in Workers adoption because so many European customers just deployed a worker that said, "If EU just don't serve that website," because US companies just didn't even wanna deal with it and didn't, didn't know how to. Corey: Yeah. I, I have thoughts on that. None of them are acceptable for, uh, public distribution at this point because I can't quite tamp down the blistering stream of profanity that winds up, uh, playing out. Uh, so a few weeks before this conference was a bunch of your releases, uh, during your birthday week. Folks have had time to play with them, and now basically talk to a number of the folks here about those enhancements. I think only a small subset of them started with, "You know what your problem is?" And then devolved into violence from there. So the question that has come up, though, do you consider Connect to be a developer conference? Dane: Well, I mean, I absolutely consider it to be a developer conference. We don't consider it a place to go announce things. We consider it a place to a conven- convenient, uh, experience, where you get to go meet peers, you get to go meet fellow Cloudflare employees, uh, you get to learn something. At least that's our, that's our hope. Like, that's why, you know, no one stood up on stage and said, "Guess what? You know, these three big announcements that we're doing," uh, that, that was very intentional Corey: So I wanna shift a little bit because something that, uh, you've talked about a number of sessions that I, uh, caught the abstracts of and, uh, poked my head in to listen to, uh, how have you folks changed how you're building and deploying things via the use of AI? Because I can talk about the insane nonsense that I build primarily in service of shitposting, but there's a difference between, "Hee-hee, if this goes really well, thousands of people will see it," whereas stuff that you do is now a statistically significant portion of the internet as a blast radius. I don't think you get to YOLO slam it into, uh, into creation quite the way that I do. Dane: About a, a little over a year and a half ago, we kinda wrote down what are, like, our, our AI principles were, and like, you know, number one and number two was no vibe coding, at least on, on the engineering side. Uh, and, and that's because, you know... And we also said, you know, if something gets deployed to production, there's no blaming AI either. So, uh, we still very much, you know, humans write code, uh, humans plan, humans do that. You know, they might use AI to assist them, but they own, uh, what they're building, what they're shipping. Uh, and, and we use AI as a tool to, uh, help accelerate that, uh, but not to replace hu- human judgment, uh, human thought, um, human creativity. Rita: That said, AI is really good at things that humans are bad at, right? So, um, you can- a human can review code and miss stuff that AI is going to catch. And so AI has been really instrumental to our code review process, and I think has actually preve- helped prevent incidents that otherwise would've happened. And similarly, yeah, I, I think kind of coming from a culture of ownership, like no vibe coding doesn't mean, uh, yeah, don't use AI. I think that one of the challenges that we're all kind of seeing now too is, um, there are things that vibe coding is really great at, like putting together a prototype, or we have a lot of non-engineering teams that really embrace vibe coding because it helps them build personal software

that otherwise, you know, we, we'd have to go get a SaaS license for something that serves, like, maybe, like, 10% of their need. Yeah. Um, so, so in some ways it's been very empowering, but I think being very clear about what's been reviewed by a person and what's going into production, uh, having human judgment also for, you know, which code does need to be thoroughly reviewed, um, because it's gonna be deployed to the edge, and it's part of, like, a very, very core service versus, like, maybe this is something completely new and experimental and that's, you know, only a package that we're giving to people to try out. Like Those are, those things are very, very different Dane: Or, or, or things that are just, uh, annoying. A, a hu- a human's not gonna find excitement over, uh, rewriting Next to, uh, be, uh, faster and safer to run, and also be able to run on Cloudflare. Uh- There are few Corey: humans, but I hear you. Yeah. Dane: But, uh, yeah, the, I mean, there was, there was one human originally, a s- a slop fork Steve. You know, he gave it the plan, he gave it the guidelines. I've been asking for that for years, and they were like, "This, this project, uh, this, this is horrible." Uh, um, and then finally, like the AI was like... Well, the AI didn't have a choice. Yeah. Rita: Well, the, the interesting thing about that is he tried that experiment twice. Yeah Once in October of last year, and he kind of kicked the tires with it, and also got bored with it- Yeah ... because the AI wasn't good enough. And then fast-forward to, um, you know, the models that came out in like November, December, and all of a sudden they're cooking with fire. Corey: Yeah. It's odd. There- there's a weird judgment aspect of it. People using these things love it. Uh, people have to consume the thing that gets put out, and the other end is awful. Like, uh, my, uh, AI factory, your slop cannon, and it, it becomes a different challenge. Something we've learned pretty clearly over the past year has been that vibe coding is awesome, and vibe maintaining is terrible, and vibe maintaining someone else's AI slop is like they're building an annex onto the last circle of hell just for that. It's terrible. So on some level you almost start to rebuild things five or six times. Where a rewrite used to be a complete anti-pattern variety of ways, and now it's starting to feel like for some things it's a best practice. Dane: I mean, we always thought of, uh, Workers as a platform where people build, uh, uh, greenfield, uh, applications. Uh, 'cause it's just a very opinionated architecture that you can get a lot of great things, uh, when you do it. But you're not gonna go port an, you know, a- an application that's been running, you know, for 10 years and move it over to us. And we were, like, kind of fine with that, 'cause still there's a lot more greenfield stuff being built. But all of a sudden, uh, we're seeing people that have these prompts that are literally- rewrite my application to run on Cloudflare Workers and, and migrate it for me. And, and all of a sudden, like, that, that process even becomes cheap. Corey: For me, the wild part that, of everything you've released this year, it is the ability to have an agent spin something up on Cloudflare without even having an account, and then okay, now you can go ahead and claim it if you decide you want to keep it. Uh, just the abuse concerns on something like that, the scale concerns you have to be dealing with, too, make that... It, it's a, it's miracle from my perspective it even got shipped, let alone stuck around. It, it's one of the, the, the fastest mean time to dopamine, uh, moments on deploying something for the world to see that I've ever heard of. Dane: Well, well, so two parts to that. One is, uh, the abuse part. We, we have a, an amazing trust and safety team, which we've, you know, had 15 years- Corey: You basically have to ... Dane: Yeah, which we've had 15 years to, uh, uh, hone that skill. Uh, so, uh, you know, there's still obviously a lot of abuse, but we are, we're, we're pretty good at, uh, uh, uh, handling it. And then, you know, honestly just the, the cost, uh, to deploy something like that on our platform is really just minimal, and then that's just 'cause how we designed it. Um, and so, you know, uh, uh, the, you know, that, that, that, uh, the abuse part is probably the bigger issue, that was the bigger concern, but the cost is not really. Rita: The interesting thing is, w- we've over the past year really invested in our, uh, dashboard experience, which as someone that still uses it, has gotten so much better, but now we're seeing so many people on X being like, "Oh, I use Cloudflare without ever having touched a dashboard. I just told my agent to go X, Y, and Z. They created the account for me and everything." And that, that's the new era that we live in, and I think that that's a requirement now. Dane: It's the worst. It's not like I started getting good at it. Rita: Yeah. Dane: And then no one cares anymore. Corey: You have to have a dashboard because a demo on stage of, "And here it's working with Claude Code," like, what, what product are we demoing here doesn't work. But the thing that works well for customers who are using it is not necessarily the thing that looks great on a stage. Those are two very different use cases from a UX perspective as it turns out, and you have to have it. It is a baseline requirement, but your common case increasingly is customers never see it. Rita: I think an interesting take on it maybe is that, um, it, it allows us to focus on our efforts on things where you do kind of have to have it maybe to explain a unique concept or to really visualize something. And like, for example, CRUD dashboards should have never been complicated in the first place, right? Um, uh, if you have a decent enough design system in place, and that part's been accelerated for our internal

teams, too, through having a design system where, um, we actually increasingly even have PMs that ship parts of the dashboard because- We already have the predefined criteria Corey: Well, that's the whole cloud company problem. To create a design system, ha, you only have one, we've got eight, and there's where the problem starts. It's... You can, you can tell when you're leaving one team's remit and heading into another. Consistency is horrible. Dane: Well, we like to pave our own way. We went 15 years without one. Yeah. Uh, and so we only have one. Corey: What I found interesting, some of this even exposed itself through CF, where there is still some inconsistency between namespaces, um, how things work with various services, and I, I still maintain fix it now because this problem only gets exponentially worse with time. But it's a... It is one of those areas that's very hard to wind up getting deployed it- because that requires centralized coordination in a company that benefits by launching things without that centralized coordination. This episode is sponsored by my own company, Duckbill. Having trouble with your AWS bill? Perhaps it's time to renegotiate a contract with them. Maybe you're just wondering how to predict what's going on in the wide world of AWS. Well, that's where Duckbill comes in to help. Remember, you can't duck the Duckbill bill, which I am reliably informed by my business partner is absolutely not our motto. To learn more, visit duckbillhq.com. Rita: It's definitely a spectrum, right? If you want innovation to go fast, then making it all go through a bottleneck is really, really hard, and so you kind of have to make the trade-off of, um, letting teams sometimes define their own. And, uh, a, a big part of our culture is ship but fix quickly, and don't let issues sit around and linger because the more they collect dust, the harder they become to fix. Dane: A- AI's kind of helped improve that somewhat. Uh, I mean, we built this thing called, uh, uh, our engineering codex, uh, which, uh, is, uh, you know, a set of RFCs on how we build. Uh, and it turns out agents are really good at following... You know, we should have had those rules already written down for humans, but we didn't really. Uh, but, uh- It's on Corey: the backlog. Yeah. We'll get to it Dane: next- Yeah, yeah ... Corey: sprint for sure, Dane: right? But it, it turns out agents are even better at, uh, following instructions when you write them down. Uh, and but then, you know, that we pair that with another kind of Cloudflare principle that we like, uh, like guidelines and not rules. Uh, so you know, uh, uh, you know, 'cause ru- rules were made to be broken, so you're allowed to ask for exceptions. But, you know, you need to have a good reason, it needs to be thought out. But, uh, uh, at least we're all starting from a common place of, uh, here- here's how you build on Cloudflare successfully. Corey: Internally, how do you handle the approach of autonomy with AI? Because on, on the one hand, just go ahead and do whatever your little heart desires. Oh, no, the payments database is not great. Despite all the beautiful lies we tell ourselves about Agile, when it comes to payments, we all do Waterfall for very good reason. Uh, on the other, you wanna go full-on cautious. Great, now I'm just hitting approve on everything. I'm getting decision fatigue, and I feel like a Claude condom, which isn't a great place to be in. Like, that, that feels like my job of hitting the Enter button could be replaced by the Dickie Bird at some point. How do you find the right place on that continuum? How do you think about it? Dane: Well, I mean, there are one-way doors and, uh, not. I mean, if you're deploying something, uh, uh, to the core runtime of, uh, workers, that needs to go through a lot of reviews. And, uh, uh, but if you're deploying a worker on that, which, uh, is gonna have a much smaller blast radius, and then you can kind of, uh, uh, manage the surface area, then you might have a different set of criteria. Corey: The problem I found is that it's very easy to identify a one-way door usually after you've gone through it. It, it... I found interesting novel failure cases of AI doing- Call it curious things. Uh, uh, no one is as zealous about backups, for example, as someone who's just lost data. Like, it feels like it's one of those everything's easy to predict after it happens. H- h- what was the process of getting there like? Dane: So we have a, uh, a, a pretty, uh, regimented incident process, uh, where, uh- Corey: Your write-ups are gold, absolutely gold, Dane: and public is- Uh, and that's, that's the kind of the, the final outcome of it. But, but, you know, it starts from when we declare it, uh, to how we document it throughout it, uh, the follow-ups, the prevents that come out of that. Uh, and that's also what led to writing the codex of actually documenting, "Okay. Well, we know what works here and doesn't work here, and here's how to build if you want something to work." That's been, uh, uh, you know, very successful for us and... But, you know, I think it's a constant lea- I mean, we're, we're learning and documenting and, and hopefully, you know, it's, it's okay to make mistake once, most mistakes and, uh, you know, we don't, we don't punish anyone for that. Uh, i- if you make a mistake twice, then, uh- Corey: It's hard to fire an agent, too- Right ... which makes it even harder. Rita: But, but that's why we have humans, right? Because some of that is very nuanced, and I think people are pretty uncomfortable with that. People want, like, very clear, sometimes rigid rules about, you know, like, "Tell me I'm allowed to b- deploy X but not Y." And it's like, well, if it was that simple, then an agent could do it. The thing that I'm asking you to do is exercise human judgment and come back to me and tell me, like, "Okay, if we released this thing on a Friday

afternoon, could I explain to a customer w- what happened if it caused an incident, or do you feel so confident that, um, actually, you know what, ship it on a Friday. Why not," right? Dane: Or it's a brand-new product and no one's using it? Rita: Yeah. Dane: Sure. Ship it, ship it on a Friday. Corey: Yeah. There's almost like a tiering ring of at some point, oh, wow, this is, this runs the core network. We should probably have some diligence around this rather than, eh, it's probably fine. You'd hope. Mm-hmm. But then you see some of the tier one ISPs and you start to wonder how accurate that really is. So I'm curious at this point of it, it seems like every time I turn around and think, well, okay, this is what Cloudflare does, I draw a circle around it in my mind. And reasonable people can disagree, but for a long time, the primary value was back when it was hard to get a domain name with an SSL cert in front of a website, great. It was the thing that I went to. Y- yes, your website talked a lot about security, but people only care about that after they've just failed to. Fun day to say that with the entire internet melting down from a fundamental security layer, but which that's neither here nor there. Then it became this idea of, uh, of being increasingly interesting things, usually at the edge. It was where you were the strongest. And now it seems that with things like object store, things like D1, things like durable objects, things that are, that are turning into an absolute reimagining of what w- a cloud provider is, where it starts, where it stops, it feels like it's very hard to figure out what you do in some ways. I think probably for me, the straw that broke my back was the idea of using you as an AI gateway, where all inference can route through you. Now it feels like, okay, I start to need like a flowchart to figure out what aspects of this you are and what you aren't. Am I the only customer you have that's starting... I guess, struggling to keep abreast of what it is that you're putting out, or am I... I mean, again, I could easily be the dumbest person in this room. It would not be the first time. Dane: No, I'd definitely say, uh, uh, one, one of our Achilles heels is, uh, uh, customers tend to not know everything we do. And, uh, you know, uh, w- we- we tend to, uh, uh, you know, have, you know, a customer that comes from this, you know, one area and they know the workers very well, or, uh, they know our bot management very well. Uh, and that's probably how we were telling the story. Um, uh, you know, uh, but I think, you know, over the past year and, uh, uh, uh, hopefully at the, uh, our, our keynote here at the, the trade show, uh, uh, you know, people really got to see, uh, how we can explain the platform in a way, uh, that is a little more coherent and that kind of explains, uh, what we do and everything it- possible to do on it that you can kind of imagine. Corey: Part of the challenge too is that what is possible on the platform today and what is possible three months from now are going to not be completely congruent. I mean, one of the hardest problems in this era that we're in is building any plan for a release or a talk in advance. I gave a s- a, uh, talk earlier this spring about how AI speak Terraform like a tourist, and that thing aged like fine milk. It's, it's one of those areas of, yeah, it's really good at that now, which is helpful because Terraform clearly hates its user. That's kind of its feature, but we all put up with it. Now we don't have to. The other piece of it though is that I still maintain that The areas that are not gonna be industrialized are wisdom and judgment. And I'm hoping that that holds true at some point, but even if it do- if it doesn't, I'm not quite sure where that leaves us. Like, at some point, like, "Oh, yeah, your job is to sit here in case the agent screws up 'cause we need somebody to sue." Uh, uh, that's not a great seat to be in. Rita: First of all, I'm glad you're talking about how hard it is for people that present and speak, that you can't reuse talks anymore. The, the, that thing becomes out of date. Like, I used to be able to reuse things, like, six months apart, a year apart, but that's no longer the case. Corey: No, they, they, they melt. They are absolutely unusable. You can reuse a few of the jokes, a few of the general themes, but well, get out, get out, get out. I start... Even when I give talks now, I feel obligated to call out the date when I say it because otherwise it's gonna turn into a, "Ha, look at this bad take you had." Yeah, what, 2004? Imagine that. Rita: It, there's kind of two things to what you were saying before, too, right? One, that's what Cloudflare has been doing has evolved and continues to evolve. And sometimes, um, yeah, the, the release cycles are, our customer... W- we've always prided ourselves in innovation, which means that we release a lot, and our customers have always struggled to keep up. Um, but I, I think that, uh, yeah. And then there's the second part of it, which is how do you communicate it? Um, which, which I, I think we're getting better and better at. But I, I think an interesting thing, too, is y- you know, you said another word that we would generally shy away from, whi- which is edge, which is where I think people kind of get confused about what we do, right? Because they associate Cloudflare very closely with, like, "Oh, I would wanna use you for things that are very close to my users." And as you were saying, we're increasingly doing things like durable objects or object storage, and I think what that plays more to is we never thought of ourselves as an edge for that reason of we think of ourselves as a smart network where we can route things to the place where it makes the most sense for it to run. It doesn't necessarily always have to mean, like, everything everywhere all the time. It can also mean In one place and at one point in

time, in another place and another point in time. That's what makes sense Corey: Now, the, the question of course is how you make that legible to people, where, huh, I have a bunch of users in one part of the world, and over time they find the experience gets better as the, the smart network, for lack of a better term, winds up relocating things closer to them. It does lead to challenges with things like data residency, which is becoming an increasing problem as various governments realize that they could legislate themselves a data center investment if they say the right things. It is a brave new world out there in a lot of respects, and it's certainly not getting simpler. I think the idea of being able to hold the entire stack in one person's head has long since ceased to be an accurate representation of reality, but I think we've held onto that fantasy for a while. It's getting harder and harder to credibly cling to that belief. Rita: I mean, I wonder if you fast-forward how many people will even have the stack in their heads, and how much of the stack is in their agents. Corey: A, a question that I'm having right now, and, and as a platform provider with a bunch of partners with booths and whatnot, many of whom fall squarely into this, what is the defining characteristic to your view of where someone should build their own implementation of a thing versus going and paying for a product that does that thing? Um, an easy example here is, great, I'm not gonna be able to go out and build a global network for, uh, for reasonable use, especially given the current price of components for things, and I don't have eight years to spend playing that game. Clearly I'll go with a provider that it's out there. But if I wanna have a, a service on a website that does a thing that's basically a, a digitized version where its biggest competitor is Microsoft Excel, well suddenly now I'm debating do I build it or do I buy it? How do you see that unfolding? Dane: So, uh, at least for Cloudflare we, we kind of think of, uh, uh, what are kind of the primary systems of record, and where is that kind of data gravity? And we wanna reduce that, but we wanna put, put it in places that have, uh, you know, that actually are pretty trustworthy software. Like, you know, uh, for HR systems, we wanna use something that un- understands compliance. We don't wanna vibe code compliance. Corey: All the Frontier labs are outsourcing payroll, for very good reason. Dane: Yeah. Things like, uh, payments or finance, uh, uh, tooling, uh, you know. Uh, I don't see, uh, any- anyone, uh, vibe coding an ERP anytime soon. That's just also because the risk, right? Like, y- you know, the CFO doesn't wanna go to jail 'cause someone, you know, decided to use a, a bad model. Um, and so, so, you know, for us th- those are the types of things that we, we've said, "Okay, well, th- those are gonna stick around 'cause risk- that, that's, there's a, kind of a risk," but we'll give people a playground, uh, kind of above that. We'll make sure that that data's accessible, easy for them to use, put kind of governance around it to make sure it's not leaking. Um, and, and let them build their own interfaces on top of it, but, uh, keep the workflow and the, uh, um, you know, kind of lo- lower level engine inside of those places, Corey: you know? Dane: How do you think about it? Rita: You know, I agree with Dane. I, you know, I think conventional wisdom has always been th- does it make your beer taste better, right? So if it doesn't actually help your customers, then, um, you probably should buy it rather than build it. Uh, I, I think that that's no longer entirely true because we do spend so much time using tools that meet, like, 30% of the thing that we needed them for, and the cost for building, yeah, especially an interface on top of a system of record is so, so low that there's no reason not to just kind of build your own. Corey: It, it feels like the open source world figured this out a generation ago. Software could be free as in puppy, where the maintenance becomes a nightmare because it's every... Or you build a thing that service- that perfectly services your needs, and as soon as you get a second user, oh, hell is other people's requirements. Now it seems like that is less of a barrier than it once was, but there is still value in folks that think very deeply about a problem set. Rita: Yeah, I mean, I think it's okay for things to just be single user now, too, right? Like, uh, I, I think that that's where the maintenance burden becomes expensive is once you get the second, third, hundredth user. But there's, there's certain software that's like, this is for me. I want a view of my team so that I can, uh, you know, do a s- a particular activity maybe at a team onsite, and I... No one else is gonna use it, and it took me five minutes to build. I didn't need to go out and, um, actually sign off all of Cloudflare's, um, you know, data f- for me to have access to a tool. Dane: Or, I mean, we're seeing, you know, every, every person in sales building their own dashboards and their own, uh, tooling on top of Salesforce, right? Which would just be unthinkable in the past. But so that mean, can come as personal software. They don't have to worry about it scaling out to the rest of the company. They just have to worry about it doing what they need it to do that day. And, uh, you know, uh, I'm pretty bullish on personal software, uh, which, you know, I think fits very well into Cloudflare 'cause you can deploy it for one person and, or deploy it for 100 people, and it kind of scales the same way. Um, and, uh, you know, and, and every once in a while one of those pieces of personal software might end up being something that- Uh, you know, it, it, it is a super hit, and it does need to scale up Corey: I, I'm also a believer increasingly in the idea of killing your pets, where all right, I've built some personal

software that I've been iterating on, and it's gotten... It's drifted a lot for its use case. Great. Throw it away and start over. Uh, here are the interface points. Maybe plan for data migration, backups, again, your friend, but build a second one in parallel. See if that winds up solving the need better. It almost becomes a schedulable tradition. Mm-hmm. Rita: I think that's where Cloudflare is not only uniquely suitable for scaling up but also for just stay- staying scaled at zero, right? It costs you nothing if you're getting no requests, and you're not doing anything with it. Corey: Uh, I guess the last big topic I want to get into with you folks before we call it an evening here is for a long time folks were talking about dead Internet theory, the sort of conspiracy theory that the Internet was largely bots talking to each other, and now that is very clearly a credible way of thinking about this. As the, as an increased percentage of traffic, I believe might be a majority now, uh, becomes, uh, automated rather than human driven or human directed, that changes the Internet foundationally. How do you think about that? Because, again, robots do many things, uh, more efficiently than humans. One thing they don't do is click on the ads, which is gonna be a problem for an awful lot of business models. Dane: Yeah, I mean, uh, you know, right now humans and, uh, agents are kind of surfing on the same, uh, Internet, and, uh, it's kind of this messy period where, uh, that, that needs to work itself out. Like, they probably should not be on the same Internet. They, you know, there probably are more efficient ways of the, each of them interfacing with it and as well as more efficient, uh, ways of monetizing. Um, you know, I, I think, uh, uh, you know, when people have ads on a page, they're, they're, they know what their CPM is. They know about how much they're gonna make, uh, you know, cost per million on, on an ad. There's no reason why an agent can't just, you know, with a micro transaction just pay that. Um, and, and that's kind of like what we're, we've been building with our monetization gateway. Uh, you know, put sites behind us, uh, let people decide what their CPM is, and, uh, uh, you know, give a, give an agent a wallet, and, you know, let them go, uh, go wild, uh, uh, using it. Uh, but, uh, we are trying to give, uh, you know, the, the site owners and creators, uh, those type of controls, whether, uh, uh, let them decide, uh, who gets in and who doesn't. Corey: Of course, defaults matter because that's going to be what most people will wind up taking. But it's a- It's weird. Some folks absolutely do not want, uh, agents interacting at all with their site. And one of the more inscrutable ones to my mind is Amazon's retail store, where they will actively block anything that looks to be bot-driven. Mm. It's like it- the bot is coming to your website to buy a thing and give you money in return for goods. Well, we're gonna build our own, because we're Amazon. It's gonna have a terrible UX, it's gonna ship three years late after everyone else wants it, and you are just gonna be surprised when it finds lukewarm success in the market, if I can Nostradamus that a second. But it feels like it's emblematic of a larger trend, where folks are clinging to the way things used to be, almost in a denial of the fact that the sea is changing. Rita: I think yes and no. If, uh, we've also seen customers shift their perspectives, and I- I think a lot of people are evolving on their thinking about, yeah, a few years ago, the world was, you know, bot versus human, and now it can increasingly be bot on behalf of human. And so, um, you- you wanna allow that traffic. And we had a whole, you know, part of our conversation that was about, um, AEO and agents buying things because there's information available. So we're also seeing businesses that are actually leaning in and trying to inversely give agents as much access as possible to their content, because they- they- they want agents to recommend them and things like that. Dane: I mean, we're early on. We saw a lot of sites that just did not want their traffic to be used for training, the personal agents coming, Dots, Muse, uh, Grokbot. Uh, you know, that's actually changing. We're actually having people that say, "Hey, how do I make sure I let this in?" Uh, 'cause they actually do want that traffic. Um, and you know, so you know, what we were trying to work with all those providers are to make sure they sign the requests, that they clearly state what their intentions are, and then the site editor can decide, uh, you know, th- I want this in or I don't want this in. You know, for the most part, you know, uh, their incentives are aligned that, you know, they want, they want them in. Corey: Yeah, it- it's a hard problem to solve for. I don't, I don't think there's any clear right answer, no matter what you do, especially at your scale. People are gonna be mad, mad about it. That, that is what the internet does. It's very good at, uh, not accepting the two very complex, complex things can both be true at the same time, and that doesn't fit in a tweet, so we're just gonna get angry about it. It's not ideal. Um, I think that we're gonna be seeing some interesting changes. So okay, Connect 2027. Let's assume it's a year from now. What's changed? Rita: Oh boy, I, even if you said connect, I don't know, January 15th, 2027 Yeah. Corey: It used to be what's gonna be different in 25 years? And like, oh, by then we're definitely gonna have quantum computing powered by cold fusion, and now 25 years later you can say the exact same things, but here we are. Some things are very clearly changing too. So I, I don't know that I would wanna put, uh, a prediction out there, which is why I'm making it your problem instead. Dane, what do you think? Dane: A, a higher majority of people will have personal agents that are, you know, are running, uh, for

them, uh, both at work and, uh, uh, at, at home. Uh, you know, I think, uh, today, uh, you know, we hear a lot about the GPU crunch. Uh, I think you'll hear more and more about the CPU crunch because harnesses and things like that need CPU. Um, you know, uh, hopefully Cloudflare's the default place where you, uh, deploy all your durable agents, um, and everyone has, uh, you know, hundreds of them making their life easier and lets them go spend, uh, uh, time on higher value things. Corey: If I had to guess, I, I think a lot of the breakthrough might very well come from folks one layer up the stack who are building on top of you folks, where it becomes like that is the thing that becomes the household brand, the recognized name in non-technical circles. And I, I'm not sure that you'd want it to be the other way of, yes, we're now going into consumer products. Uh, feels like it's a bit of a pivot there, but being able to be the, the, the underlying infrastructure providence that everyone else starts to build upon. Uh, increasingly I am seeing at least on the, uh, at a lot of the transport layer that Cloudflare has well gotten into that point. Rita: Yeah, I mean, we were just, uh, just before this I was in a session where, um, John from OpenAI was showing how, uh, GPT Sites is built on Cloudflare. We had, uh, Fabian on stage with me yesterday who was showing how Lovable builds on Cloudflare. So, um, in that sense, we've always viewed ourselves as the infrastructure layer, and, um, our, our goal is to give people that are very good at building these consumer experiences, and that's why I think there's a lot that's been left to imagination still. There's so much that's been unexplored. Um, one thing that I'm actually very bullish on is, uh, real time and, and voice feels very early on, and we're, we're seeing more and more folks ex- experiment with that. So I, I think you're right that we're, we're trying to play at kind of the, the layer slightly below. Dane: W- I mean, at, at our heart, we're an infrastructure company. Uh, and so, uh, you know, uh, while, uh, there was a lot of, uh, you know, a lot of c- like, a lot of questions early on, "Well, why, why don't you build a, build a, you know, vibe coding, uh, website builder?" Uh, we're like, "Well, 'cause we wanna build the platform, and we wanna build the primitives that, uh, let, uh, there be 10 of them on us, um, or hundreds of them on us." Um, and I, and I think largely that's kind of played out. Corey: I, there, I feel like there are two kinds of infrastructure company in the world, the kind that are, have not... "We're an infrastructure company." Great. That empowers an awful lot of things. Please build on top of us. And infrastructure companies that are mad about it and like to delude themselves into thinking, "Nope, nope, we're gonna move up the stack now and compete with all of the people that build on top of us." I almost wonder at some level if that's an inevitable consequence

of getting to a certain point of scale, but it's refreshing to hear that you folks are not inflict- that affected by that yet. That's the one thing I hope doesn't change. Rita: We have a lot of conversations about it internally every once in a while, and, you know, what, what that layer on top keep, keeps changing. And, and one thing that we will do sometimes, um, mostly just so that we have experience in-house in understanding our customer's perspective is we will build, like, an internal ins- instance, and, um, sometimes we'll open source it as an example, right? To even show people what putting together the pieces would look like. Because also just being infrastructure that doesn't show you, you know, it's like how to draw a horse, and step one is very simple. And then, no, I s- I swear, you, you could get to a horse. Um, so, so that's one direction that we've found ourselves leaning in, but it's not to say that I, I think as a company, you can definitely build new muscles, and there are muscles that I'm very proud that we've built over time. But also understanding where your strengths are and where you're better off, um, you know, either letting people build on top of you or partnering, um, allows you to focus your energy on the things... Like, building infrastructure is really hard, and we have no shortage of problems to solve over there. Corey: You, you would think that would be more broadly accepted than it is, but now we're into the realm of people promising data centers in three months, and they think it's basically, uh, some tents and a generator. Um, good luck. Have fun with that. It, it takes work, legitimate hard work, to get these things built in a way that's sustainable and scalable, and it's nice to see that folks are sweating those details. Rita: That's our job. Yeah. Corey: I wanna thank you both for taking the time to speak with me. If people want to learn more, probably not the people here in this room, uh, you, you probably should know where to go to figure out more what Cloudflare is up to. But for folks listening, where's the best way to get started with Cloudflare these days? Dane: Go to cloudflare.com or point your agent at it and- There we go ... and ask, uh, what, what, you know, what, what can I, what can I do with Cloudflare? How can I use it? How can I make my apps better? Corey: How can I make it hysterical and shit post-y is usually my prompt, but it works. Dane: Yeah. Rita: Someone had a great suggestion the other day. We, uh, published a blog post about what, what we use for our own internal AI review tool, and they were like, "I just started pointing my agent at Cloudflare blog posts and being like, 'Extract the parts that I could reuse from this,' and, like, start implementing it." So yeah. Corey: I still have this weird behavior every once in a while where an agent will look at this, re- well, I'll point it to one of your blog posts, and it'll be skeptical and try to fact check the things that you say in the blog post. Like I, like I made it up or something, like I'm trying to trick it. I'm sure someone out there has a good reason for it, but yeah, Rita: it- I promise we do a lot of fact-checking before any of our blog posts go live. Dane: A lot of lawyers. Corey: Yeah. Yeah, it turns out that they have opinions that you, you don't heed at your own peril. I've learned that over the years. Thank you both so much for taking the time out of your day to speak with me. Thanks again. Thanks. Dane: Have a great one. Thank you. Thank you.