npm install Agents — with Sunil Pai and Rita Kozlov (VP AI) of Cloudflare

So welcome to the Latent Space Lightning Pod. Uh, we like to do these as sort of quick reaction things to things that have just been launched that people wanna know more about, and I think Cloudflare launching an agents product is definitely up there on the list.

Uh, I'm super, super happy and honored that, uh, I've known both of you, uh, since before you were famous, uh, maybe, maybe a little bit after you were famous. Uh, so welcome Rita and Sunil.

Hi. Hey. Hi, nice to see you, Shawn. Super happy to be here.

How do you like to introduce yourself? What's your, what's your quick bio that, that you wanna give people?

Rita first.

Um, been at Cloudflare at n- for nine years, and, uh, lead product for all developer things Cloudflare.

Been at Cloudflare for a total of two-ish years with a little bit of a gap in the middle to build my own startup. I became friends with Rita when Durable Objects were announced, and I was four drinks down and tweeting at her, asking her to give me access, and then I started dissing the developer experience, so they offered me a job.

Yeah, if you're-

So many great things do come out of the random Twitter interactions.

Yes.

And Cloudflare employees are, like, extremely online. Like in, like, it's a problem. Yeah, I think it's where developers are, right? Like, if you care about reaching out to developers and hearing their pain points, like, yeah, you know, it's, it's, it's gonna be Twitter.

So y- I'll get right to the point. There's been Cloudflare AI offerings for a while. Like, you know, I think I was pretty excited about AI Gateway, and I think what people don't really know is, like, the, the, the power of sort of edge models that are sort of run in the Cloudflare style.

I think it's still pretty underrated, and I think it will become more powerful as, as more, more edge models become more powerful. But for the first time, you're actually using a Durable Objects extraction to do some kind of agents thing.

So you launched this, I think, like, two days ago. Um, yeah, what is it? Give us a high-level overview.

You go.

Okay. Um, Durable Objects have been around in Cloudflare for about four years now, and I think they are the world's first implementation of the actor model in infrastructure, the thing that Erlang Elixir made, like, super popular. Cloudflare got that out for JavaScript developers.

And it turns out that these little instances of compute that you can write a code, that you can spin up really quickly and scales horizontally across the planet, is the perfect little container for running your agents in. So that's kind of the take that we are doing here, which is not letting you bring your own code, whether it's written in the AI SDK, LangChain, et cetera, et cetera, and we give you a container for running millions of these that you can coordinate, you can talk to with HTTP, WebSockets, email.

We're going to do audio/video also, like, just all of that. That's the take on this. Like, I think we are mo- doing the thing that Cloudflare does, which gives you the platform for running these things at scale and in a very nice way.

And the moment we figured it out last year was like a lightning bolt for us. The way Rita says it is, we have... It's not really lucked out, but, like, we have magically created the perfect infrastructure for running these things.

Yeah.

Anything, anything to add?

Uh, the way that I like to think about agents is, a- an agent is really like LLMs and a bunch of workflows and coordination and orchestration and then extra, like, some sort of services, right? Like, you have kind of the brain of the operation, which is the LLM, and it can come up with a plan, um, and then, but then it needs its executive arm kind of, right, to, like, go and make sure that things happen and that even if failures occur, that, like, something is communicating back to it.

And this requires these, like, two bits of compute and state to be really, really closely tied together. Yeah, that, that is kind of at its core what a Durable object is.

Yeah. Uh, I mean, it's still, still unmatched as a, an extra-extraction. Uh, honestly surprising that, you know, uh, it hasn't really been cloned, uh, successfully anywhere else. But yeah, I mean, uh, I'm curious to see... You know, you, you hinted a little bit in our pre-chat about there was a sort of fun origin story of this agents work.

Why not go into that a little bit? How, like, how did, how did this sort of focus come about? Like, Durable Obje- Durable Objects has, has been there for four years. Uh, what is this recent push to make a proper Agents SDK?

Well, there are people who are already building agent platforms on Cloudflare. I, as you know, have been a big fan of Durable Objects for years now. That was the big focus of the startup. And I even wanted to do, like, agents then, like back in like '23, but it wasn't clear then, like, what the story would be.

A few months ago, we started talking about it internally, and we realized that there was, was clearly a fit. And I think we had the option of saying, "Oh, are we going to build, like, a no-code platform? Are we going to do, build another library that campaign, c- competes with, like, libraries?"

And, uh, I think, I think I was talking to Ben at the time, my manager, and he was like, "Why don't you do what you usually do, which is, like, just go out and, like, talk about it, like, just literally build in public," et cetera, et cetera.

Vibe coding.

Then two weeks ago, two weeks ago, two weeks ago, I built a shitty first version, and I to- I mentioned in the chat room that I'm going to publish it. At which point, Rita stepped in and said very politely, "Hey, this is a great start, and let's just make it better.

Let's actually, like, announce it properly, like put the effort." So the whole organization came behind it, and we did like a big crunch to make it actually be good. That's, that's kind of what happened, which is they prevented me from embarrassing Cloudflare globally.

I mean, we're always down to ship early, and I, I think Sunil is kind of actually such a great example of someone that will shamelessly build in public and get so much great feedback along the way. But we're like, "Okay, i- if we're gonna do this," especially, I, I think one of the things that...

We've been thinking about this for a long time, and the thing that we don't want to do is clutter the ecosystem. I, I feel like it's almost a running joke right now that everyone is building an agents framework.

So we were like- Do we, don't we? Do the current tools solve the real problems? And so what we didn't want to do, actually, is create something super heavy-handed that feels not yet where we need to go. But we're like, "Okay, we have the kind of right underlying primitives for building this," but there, there's still at least one thread a week that goes something like this: "Hey, can someone describe durable objects?"

And then there are 100 comments, and no two of them are the same, and people walk away equally confused. So we're like, "Okay, let's help bridge that gap and, and bring people there." We started, uh, and Sunil kind of took a really good first stab at what would this look like and, um, actually connecting it with, uh, you know, the front end and the React ecosystem and all that.

I, I think we're instantly like, "Wait, this is slick. Let's, let's do this."

Yeah. I think, I think there's a lot of synergy here. I, I think this is a good point to where we should start showing some code. Uh, if, Sunil, you wanna, uh-

Sure. I can just jump in

... start figuring out the screen share, uh, thing that is useful.

Let me share my screen.

Because, like, yeah, that's, that's where, that's where I think, like, people, uh, fall apart, which is, like, not sticking enough to, like, okay, like, how do you use it? What does it do? What d- doesn't it do? Yeah, go ahead.

So first, a couple of links. Of course, all the code is under Cloudflare/agents. We have a whole bunch of documentation under developers.cloudflare.com/agents, and this we are updating, like, on a daily basis. Shout-out to Matt. Like, he's been, like, grinding on this, like, a lot.

And we have a starter kit at, under Cloudflare/agents-starter, which is, as is the style in JavaScript development, there's a little one-liner that you can take, run it in your terminal, and it spins up a project that has not just the agent framework, but a little front end for it that has React and a little chat agent starter

that lets you just take it, and, like, you can basically start adding tools into it and working with it. So let's just go into the code. Let's see what it looks like.

As I mentioned, so there's a bunch of configuration stuff because it's not a JavaScript project if-

This is-

... there's not 12 configuration files in it.

That's something I always ask. When you screen share, can you bump up the font size?

I... Of course, I can. Yes. Please see all my configuration files. Nothing out of the ordinary, to be honest. Like, it's, it's wheat and a bunch of TypeScript stuff. We use the AI SDK for, uh, the starter kit because it's actually pretty good.

Okay. But if you wanna look at what an agent looks like, here we have an example of a chat agent, which is just the most common one, and I've been converted. I used to think that it's for amateur use cases, but I actually think it's extremely powerful now.

And the way that the AI chat agent class works that we have exported, and we'll have a number of other agents in the future, but this is the one we have right now, is you literally just tell it what to do every time a chat message comes in, and here we have implemented, hey, like, it, uh, takes tool calls, does stream text, have a little bit of a prompt here that does, uh, this thing, and it returns it.

And every time, uh... And this we have for... I should probably call this execute scheduled task. How d- what happens when a scheduled task is triggered? What is a scheduled task? Let's just start this up and see what it even looks like, what the starter looks like.

Fantastic. Uh, as you can see,

it's a regular chatbot, uh, and you can say hi, and I wired it up to OpenAI here, and, uh, it responds like it would. Uh, there are, like, two or three tool calls that we have in here. So for example, if you say, "What's the date?"

It says it's Feb 27, 2025. Now, we have also implemented human in the loop, and we like- We've made all these fancy terms extremely boring. We have a hu- human in the loop, uh, confirmation step if you ask what the weather's like.

So you say, "What's the weather like in London?" And it i- immediately asks you for a human in the loop step. But the nice thing about this, first of all, you can see that the messages actually persist. I'm refreshing the page a couple of times here.

So much so that even if I kill the server, refresh this, come back, I come back, it, uh, still stays there. So you can approve it, and it goes back. The nice thing about these steps is that they can happen in the background.

That's the specialty of using a durable object for this, which is if you clo- sh- if you shut down your laptop or whatever, it's fine. It doesn't need any open requests. It's all ru- actually running on an instance of a durable object or agent instance that's conveniently, it's spun up about 20 milliseconds away from me once it's in production.

What's also nice is that we have scheduling support built into all our agents. Uh, what does that look like? You can s- ask it stuff like, "Hey, uh, every Friday evening, remind me to call my mom." We have a little thing that takes that, and it extracts it, and it does, it schedules a task to run there.

What's nice is that it doesn't just take dates. You can even give it, like, patterns. Uh, well, this is a cron pattern. It literally says, like, every Friday evening. But you can also say stuff like, "In 10 seconds, tell me a haiku."

Ooh. It's like-

It actually schedules a task to run in 10 seconds. You wait one, two, three, four, five, six. I should have made it five seconds, I guess.

No, but it basically, it's gonna... Yeah. Oh, oh, it's go- but it came in as a user.

Oh, it came in as a user me- okay, fine. I need to-

That's a, that's a little bit of a, a hack. Yeah

... fix my 06. Uh, but point being that even if you're, e- even if you weren't connected, it would, uh, it would trigger in the background. And over the next couple of days, I'm going to add email support for this.

So if it notices you're not even connected to your chat, then it'll send you an email, and you can respond to the email, and it'll update the chat. So we want to- want to think about multimodal not as just something that's a web or audio video, but, like, we think email is super important, and you can chain these tasks together.

You can basically say, "Can you tell me what the weather in London is in 10 seconds?" Et cetera, et cetera. And that's where the agentic behavior starts coming in. After a point, you can have things that run fully on schedules.

Any other system messages, anytime a user signs up, send an email. But we think this is an interesting starting point for running these things. Does that help?

Yeah. Anything you'd add, Rita, uh, in terms of, like, what you see here from your eyes as a PM?

Yeah. I, I think as Sunil mentioned, I think the how do people call agents or interact with agents, the entry point story there is really interesting,

which is where I'm really excited about email. What else would I add? I don't know. I'm actually kind of curious your perspective for seeing this for the first time. What stands out to you as useful? Um, I, I think one of the hard...

And where I'm actually kind of curious your take, I asked myself this question the other day, and then I asked it on Twitter, and it didn't get very many responses. But one of the things that I think is telling of where we are in the agents conversation more broadly is there's not really a canonical hello world for an agent.

Like, I feel like the hello world of LLMs is RAG, right? Like, everyone has built a RAG at this point, and we're at this interesting kind of uncanny valley with agents where we don't... It's, I think the hard thing about coming up with it is there's nothing that you trust it enough to do that you're like, "I'll give it the keys to the car, and I'm fine with it running off and doing," right?

So, like, something that feels very agentic to me, for example, is, like, okay, don't just suggest restaurants for me. Go make me a reservation. But then, you know, I need to start giving it a lot of credentials and my credit card information.

And so I think that that's kind of the next interesting bits to work out and, you know, where human in the loop and all of that is important because we're still gonna be very reliant on that. Like, "Hey, run everything by me for a second until I really genuinely trust it."

Yeah. Um, I do think those kinds of escalation of authority or, like, defer to authority is important. Yeah, I think, you know, like, people are still learning about, like, the, the ideal abstractions for, for this stuff, and making it easier for people to build is, is important.

I definitely think, like, the actor model for an agent, uh, makes a ton of sense, and I think if people haven't come across the actor model, haven't really internalized what it means yet, they're, like, it's completely new to them because it doesn't exist as a thing in most programming languages unless you, you know, you kind of really treat o-object-oriented design as, as, like, a, as, as a first-class citizen.

Yeah, I mean, like, I mean, uh, this... I was reflecting that Sunil basically built ChatGPT Tasks for fun, and it's easy because of Cloudflare's infra.

Exactly. We're probably going to backport that API into, like, regular durable objects now, like, out of the library. It's just so much fun to use. Uh, I also want to, like, go back to what you mentioned about, like, understanding what the actor model is, et cetera, et cetera.

Well, this is a good time for me to shill that there's actually a lack of, there's a lack of writing and information around the boring parts of the actor model or even, like, AI agents. Everyone talks about it in a very, like, elevated manner, but we are taking the boring parts.

We're like, "No, this is, like, for business automation, et cetera, et cetera." So please visit my blog. I feel like I'm a teenager again telling people to go to my blog. But I've been doing a little bit of writing on this trying to figure out, like, what the vocabulary for it is, just the low-level stuff.

And talking about it in the context of AI agents is interesting because I don't have to use the phrase durable objects many times then.

Yeah. Yeah, yeah. Um, I, I agree with that. Uh, I noticed that you had a thing on local-first. What is the Cloudflare stance on local-first stuff?

I have an opinion, but Rita, do you have one?

Like, do we have a stance?

So for those, for those who may not be familiar with the context, well, I'll bring in my context, and you can, you can correct me. There's this movement that's kind of building, mostly led by the Ink and Switch, that it's basically trying to be linear for everybody, right?

It's the, it's the shorthand, right? Like, uh, write to a local copy of your database first and then have it sync up to the cloud instead of writing, uh, calling APIs and waiting for it to respond. That is always slower because of the network hop.

My impression is that Cloudflare kind of eliminates the need for that because durable objects make sure that your database is located, you know, to the closest point of presence. But it is still not local-first. Do we care about that?

So I have two perspectives on this. One is in a world of applications which are moving away from static content into live context-driven content, very dynamic, very highly interactive stuff, technology like WebSockets is just standing to win. And in that world, having...

So the core thing about local, local-first is also about, like, owning your own data and stuff, which is interesting, but I think the sync engine is the most interesting part of it. And Cloudflare has very unique... I happen to think we do WebSockets the best just out of the box.

I mean, like, I built a company on that, and it worked out, like, really well for me. But the other thing that you mentioned about durable objects kind of eliminating the need for local-first, you don't want to look at it as eliminating the need, but you can think about it as instead of the browser being where you sync the data, you can have a durable object that's syncing your data.

Imagine if the durable object itself is a local-first client that is pulling in data from multiple sources and live updating. And the interesting thing about that is that it can be doing that even when you're not on a web page or open up the app, which is what-- which is a trade-off that you have with local-first systems right now, which is you go on holiday for a month, and you come back, and you open up chat, and you have to wait for five minutes while all your messages try- stream in, et cetera.

Nope, your durable object has been syncing all that time, and you come back, and it's instantly available for you. I hate to be the guy who's like, "Oh, durable objects solve so many problems," but bro, like, it solves so many problems.

I mean, that's been our-

I, I think that's the state. Huh? Sorry?

I, I think that, that, that, that's kind of been our hypothesis for a long time, I would say, like, before local-first as a movement or an expression kind of existed as, uh... Previously, there existed-- I, I think a lot of the stuff stems from historically there being two places where data could exist or compute could exist, which was device and centralized cloud.

And when those are the two choices, then local-first makes a lot of sense because, yeah, you're paying that massive network hop every time. When you can-- Like, I think we view ourselves as very much best of both worlds, where you can get so close to the user that, um, that it becomes, I don't wanna say a little bit more moot, but also, in a way, it's actually kind of easier to build local-first applications, um, b-because y-you're able to...

Like, actually, the Durable Objects model is really great for that because it is designed around, like, a per user use case, and you can hold that user state, and it will get created closest to where you are as opposed to so many other, um, sync systems or, you know, traditional, uh, data systems where all of the data is centralized around where most users are.

Yeah, I think that-

Nice. Yeah. Go ahead.

Sorry, you were muted. Go. Yeah, um, yeah, I think, like, that, that makes a ton of sense for me, and I, I think, uh, these are infrastructural things that basically, like, there's almost like a, um, a-- The, the Sapir-Whorf hypothesis, right?

Te-says, says like, you know, your language shapes your ideas. If you, if you don't have the language to express it, and then you don't have, uh, you don't have the ideas that come up. And I think, like, over here, it's kind of like your infrastructure shapes your, your apps.

If your a-i-infrastructure doesn't make it easy for you, then you just won't do it. So-

Hundred percent

... yeah, be passionate about that. Okay, cool. Like, um, anything else that, like, you guys think, uh, are, you're focused on or interested by that, like, are, are trends that you, that you are, you know, is top of mind for you to, to build an agents that as well?

Maybe I'll give a shout for something that I think is really interesting and we're actively soliciting feedback on. We kind of started with the, how do we help you connect or the orchestration piece of agents. But talking to people who are actually building agents out there, I think the next challenge is observability.

And, um, okay, I've gone and spun off. I have hundreds, thousands, maybe millions of agents that are running. How do I track their different states? How do I track where they're stuck? I, I think that that's a really interesting next problem to solve and something that we're actively thinking about.

And so if anyone out there is also thinking about this or building an agent and running into these problems, I would love to talk to you.

Sunil?

Uh, like I said, I'm focused on the boring part of how do we get this into the hands of normies. Like, I was just talking on Twitter about this today. Like, how do we get the hospitality industry to use this?

Like, and not in ten years, but in, like, within the next six to nine months. Because that's where the real application, like, really starts affecting the economy or things like that. Which is why, uh, th-which is why I keep s-saying the phrase business process automation, but also why I want to see some boring progress in multimodal audio-video.

How do you-- Like, w-what is the story for tool calls at the app layer, but with audio-video? It's, like, such a hard problem to even, like, hold in the head. Yeah, I, I... Like, we need to convert this technology into something that's about AI can fuming to, like, nine-to-five developers in, like, JPMorgan.

Like, that's not even JPMorgan, but really, like, mom and pop shops, and I'm very, very interested in that journey. Like, getting this into the hands of regular-ass people, I think is going to be a big deal.

Well, uh-

Yeah, uh, sure. My thing's almost like government services. You, you want the highest, single highest leverage point. Let me file my taxes using this stuff. Let me look, look for my deal with the DMV using this stuff. It should, it should all be there and...

Sorry, go ahead, Rita.

No, no, no. Uh, I, I think you're exactly right. The, the DMV appointment is the one that haunts me. That's when I'm like, "I, I know the agents are real when I can do that." Your normies comment, Sunil, also made me think of another thing that's very top of mind for us, which I'm very bullish on MCP.

I'm really, really excited about it. But I think funnily enough, the term MCP is the thing that scares the normies away because it sounds so deeply technical, even though really it's the part that I think is going to make agents interesting because it's the integration to, you know, again, going back to the reservation example, right?

Like, how do I make it really, really easy for, for someone to build something that does that? And I think that we're gonna see a wave of ho- you know, going back to, again, Sunil, your point about mom and pop shops.

These businesses turn up that are agent first, and that's kind of the only interface to using them, as opposed to, you know, UIs and APIs and other things. Just, like, connect your AI to it and it will do XYZ on your behalf.

Yep, very cool. Cool. Uh, the, uh, you know, I think, I think this is really great work on the evangelizing agents. One more, one more thing I, I wanted to reflect on is that Cloudflare, well, you have a very sort of dev-first culture, and I think definitely the, the workers team seems to be like that.

But, like, I, I, I think one thing I, I don't hear enough about is, like, sort of like what your customers are telling you that, that they want. Uh, it's almost like they're, you're sort of building out the capabilities and then waiting, waiting for the customers to, like, come in.

Uh, like, uh, you know, o-one thing I, I would definitely wanna see is, like, how people end up using the agents. Obviously, you just launched it, so it, it's hard, it's really hard to know. Um, yeah, like, this is one of those things where I think there'll be a million, uh, shots at it, and I think Cloudflare, for, for whatever it's worth, you know, has one of the best infra, uh, products, product suites, I guess, together.

Um, and, like, people have to pick their choices, I guess. Um, I, I don't really, I don't really have a perspective beyond that.

Yeah.

Is there anything else that you wa- you wanna ask me, actually?

Have you built an agent yet?

So it really depends the definition of an agent, right? So by many people's definition, like, a agent needs to have a side effect on the world, therefore they have to read and write, therefore read-only agents don't count. Right.

So in, in that sense, no, I have not. But I think a lot of people would consider deep research an agent, even though it only reads. Uh, and in that sense, the, the AI News product that I have is a whole bunch of agents ju- that just research things.

Yeah, I, I think that's the maturity process, right? Like, it, th- it's not a binary of, like, what is or is not agent. There's just a spectrum of agent-ness, and the more things you, you throw onto it, I think the better.

The more IO formats that you put... you do, like, you know, email or voice or whatever, the better. And, uh, yeah, I think we'll all converge. I do, I do feel like, I do feel that there is not enough work on, like, making it easy.

I, I think, like, the, you know, what you've just demoed is one of the most, most advanced things I've seen so far, especially in terms of, like, how easy it is to deploy. Like, it already is running, and, like, we know that Cloudflare is, like, very good at infra, so it's gonna stay up, which is not a guarantee for a lot of things.

And, and yeah, I think... I, but I still think it can, it can be better, right? Like, um, work, like, you build workflows on top of durable objects and, you know, and you know that I have extensive history in workflows, and, like, I think there's, there's still some definitions or features of workflows that are still not existing in the Cloudflare products.

And, like, all these things need to be more mature in order for people to interact with these actor-model agents the way that we want to, which is, like, as though they were another human, basically. Like, that's the bar eventually.

See, bro, the first step was making sure you don't have to use Kubernetes, and that we kind of solved. Now we'll do the rest. Like, that's just it.

Yeah.

Like, you can, you can run agents without Kubernetes or VMs.

A huge buzzword this year is ambient agents, agents that work in the background. So you, you've demonstrated Cron. And I think Cron is, like, the starting building block of something that is more workflow-y that looks like ambient agents.

And, you know, I think the, the bar is... We actually had, uh, someone, Mahmood from Neon, teach a workshop on how to build an email agent. Everyone wants the magic, like, e- read my emails and see what's important agent.

And so technically you can build that with Cloudflare's agent, so with what you just showed. Technically you can, but you wouldn't. And then really introspecting as to why, like, how much work goes on, and, like, how do you split up that work?

How do you, how do you, like, you know, fan out? How do you store things? I think those will really introduce infrastructure insights that I can't really... I don't even have a name for it right now. It's just like th- there seems to be a divide between, like, the, the very lightweight serverless things that you guys do and the heavy, heavy work that leads people to reach for Kubernetes.

I, I mean, there's, there's the CPU heavy workloads, but I have been actually thinking about it. Conveniently, I just blogged about it today. The idea of an event bus, I think that starts becoming very central to how you think about not just organizations, but even, like, personal lives, and being able to, like, react to those systems.

Not just being able to react to events, but then, like, add more events onto the system. Cron is a capability, but... And I don't, uh, to be hon- Like, I think of Cron, like, that's cool, but I really want to get to a place where we have an event bus of events.

And then, and I really don't like the word, but it's a thing that everyone calls it, what we call orchestration, which you call, like, fan-in, fan-out. I think the LangChain people just put out a paper today which talks about the six patterns, and I badly-

What?

... want to implement all of them, like, in the next few days on Cloudflare Agents. Yeah, they, they basically did, uh, what they're talking about, the six patterns of-

Is this the Anthropic thing?

Not Anthropic.

Wow.

That's the five patterns of routing. This is the new multi-agent six pattern paper that just came out. I l- I love these people who are trying to claim Mindshare for it, because they come out with examples, and I tell Cursor, "Hey, can you rewrite this in JavaScript for me?"

I just love it.

Yeah. Yeah, yeah. That makes sense. Yeah, like, there'll be a lot of interest in that. Um, yeah, the... Actually, it'd be cool for you, you guys to figure out agent-to-agent, uh, communication. I think that is something that OpenAI is working very much on.

Like, yeah, the, the next frontier of research is multi-agents stuff. Very cool. Well, thanks so much for jumping on. Uh, I, I know it's, like, early days still, but I always want to be the first to, like, tell people about, you know, cool new things and that are backed by credible teams.

Um, and I think you guys definitely hit that bar. So yeah, thanks for joining me.

Quick note, in five weeks we have Developer Week. It's, uh, one of the biggest weeks of Cloudflare for announcements, all the fun stuff at May. Everyone gets together. It's absolutely the most fun thing to do. Watch out for that.

Where do people sign up?

April 7th through 11th.

Yeah. Where do people sign up? Like, how do you...

It'll be on Twitter. It's free. It'll be on our blog.

Follow us on Twitter.

It's one of those things, right, like, every day it'll just be on Hacker News anyway, so, like, you don't even need to subscribe to anything. You just, you'll hear about it.

Yeah.

Also true.

Although, I will say there's been an incredible thing that's happened. We recently started using a changelog as a way of announcing things that we're doing, uh, as opposed to just our blog. And so every week has kind of turned into a launch week.

We've realized that just this week, yeah, we shipped the agents framework. We shipped yesterday Guardrails, which is basically Lama Guard as a part of our AI gateway. And then today, I don't know if the changelog has come out yet, but I will share a little spoiler since I, I know it's gonna happen at some point today.

W- we're shipping a, a direct REST API to browser rendering. Um-

We sh- are we shipping that today? Awesome. Okay. Yes.

That's happening today. Uh, so we, we have a browser rendering, uh, API product that's previously been available through a bindings. One request that we would get often with workers would be, "I wanna spin up something like Puppeteer and be able to interact with a browser."

And so we shipped it for workers first, but, uh, especially with people building agents, I think it's kind of funny because it's a bit, to me, the anthropomorphization of agents, where it's like we humans use browsers, so my agent is gonna use a browser.

But eventually I think that will get replaced with MCP, but that's where we are now. And so I think that that would be a really, really cool tool for people to start using.

Yeah. Awesome. Um, it's funny because, uh, so I know, uh, we, we just did an interview with BrowserBase, and I know there's a lot of people, like basically most of the browser-using agent tooling industry is, like, I think, like, a third of their efforts is getting around Cloudflare CAPTCHAs.

So so if Cloudflare just shifts your thing, then, like, okay, like, y- you know, we don't have to spend any effort on that. You're ded-

Well, we, we wanna be honest to the sites that are on us, right? So just because you use us doesn't mean that we're just gonna let you bypass things, but actually this is a really, really big area of research and investment for us is how do we authorize and authenticate users to act on humans' behalfs.

And it's become a really... It's a really interesting problem because obviously people want to protect their properties from bots and crawlers, and there's, like, an entire category of those use cases. And agents kind of throw a wrench at this whole thing because it's, it's kind of this thing that's in the middle.

It's not a human, but it's something that's acting on behalf of a human. Um, and so o- our, our view on it is less, "Oh, we can cheat because these sites are already on us," and more, "Let's solve this problem properly for ourselves and also everyone else."

This is like the Cloudflare DNA problem. Like, solving this is, like, the Cloudflare problem to solve. Like, internally, there are so many threads about this, and you have the smartest people on, about this problem on that thread. And it is so distracting because I don't understand half the things they're saying.

I, I mean, I'm a front-end guy, right? But it's super fascinating to see, like, how Cloudflare thinks about it. I'm, like, very curious to see what we put out.

I mean, my, my hot take if, if that, if it, if it helps at all or you wanna respond to it, is that I, I'm not sure it can be solved. I think you've always been able to spoof, um, you know, user agents.

Um, you'll always be able to spoof IPs or whatever, right? Like, they just, just cannot be solved on a rendering level unless you wanna change how HTTP works. I don't know. Like, you know, like, this, the genie is already out of the bottle.

I don't think it, it can be solved there. I think the, the main thing I've been advocating for people is when you issue API keys, do not assume that the, the, the API key is, uh, is meant for a direct user.

Sometimes the user may be a, may be, uh, delegated to an agent. So in the, in the same way that, you know, every Twitter, every big Twitter account has an alt, every big Insta person has a Finsta, I think every API key has, like, an agent shadow for that API key that, that has sort of limited restrictions or at least, like, an understanding that the rate limits would apply differently.

Yeah. I don't know if-

I, I think this problem is gonna have to be solved for agents to take off, quite honestly. I, I mean, I, I see what you're saying, but we do have to at some point then pass identity onto... Uh, like, if we go back to what we were talking about of, you know, I would love an agent to be able to make me an appointment at the DMV, at some point it has to vouch that it is I, Rita, that's making this appointment.

And so that's where... Anyway, you've said two things on the podcast so far that my mind went, "Challenge accepted." Um, so agent, like, i- identification and, um, the, I think the second one was the Kubernetes problem or-

The, the heavy workload problem. Yeah, like where you need a file, you need a file system, you need media processing.

Well, Rita, should we ship a container platform?

Interesting. Yeah, you guys have been teasing that, which would be very exciting. I still remember, like, I was at AWS when we shipped, uh, you know, like, container, serverless containers, and everyone was like, "What the hell does that mean?"

But I think, yeah, we're figuring out how to cut down containers to, to make this possible, and, uh, it'd be, it'd be amazing if, if Cloudflare figured it out.

Well, the, the thing that we hinted at, I, I mean, we are working on it and, and I think that the big unlock is gonna be it's not either/or, it's a both. And the thing that we're really excited about, um, that maybe is the thing that I'll leave a teaser for is, uh, durable object enabled containers.

Uh, yeah, uh, I'll let that trickle through the brain for a second. And I, I think that that actually lends itself to these agentic workflows really well, uh, especially because I think it's easy to conflate long-running with CPU intensive, and these are kind of not the same thing.

And actually, a lot of specifically agentic workflows, a lot of what you're doing is you're waiting. You're waiting on human in the loop. You're waiting on LLM. You're waiting on random external resources, right? Like, that's not time that you're actively doing something.

And so having patterns for this stuff actually outside of Kubernetes, I think, is gonna be really, really important for doing this at scale and doing this in a way that's cost effective.

Awesome. Um, cool. Well, thank you so much. This is a really awesome overview and catch-up, honestly, between, uh, the two of you. I always treasure every chance I get to, to chat with you.

Likewise, Shawn. It's been nice seeing you. Thanks for having us.