Cloudflare’s Agents Week: Building Infrastructure for AI Agents

Hello everyone and welcome to This Week In Net. This is the special April 2026 Agents Week edition. We have a special episode about our week full of announcements. One of our Innovation Weeks. It's the first of its kind. So the first Agents Week. I'm your host João Tomé based in Lisbon, Portugal. And with me I have again for the second time in the show Ming and Anni. Hello. Welcome. Hello again. So excited to be back. Hello again. And I bet you're really tired because Innovation Weeks are all about deadlines and shipping and sharing with the world what we do, but also really busy. So how busy and full are you? I feel like last week was really the roadshow moment. You know, we did a lot of kind of talking about all the cool stuff that was coming out with a bunch of internal teams, some external folks. And then this week, I think, at least for me, it was more about making sure that the thing that we were shipping got over the line because one of the announcements was related to AI Gateway. One of your products. I think this week was really interesting as well. I feel like we still had announcements being slotted in as this week was coming along. And so we were trying to figure out the best day to position them. But once we kind of have the general story set in stone and figure out when to go out, a lot of our blog authors and, like, the teams that work on it, they're super responsible and super timely. So we were able to get everything in on time. So no worries there. That's good. And last-minute blogs, last-minute things that we ship, those are always the case for Innovation Weeks. And we're recording on a Friday, and this episode will come out on Monday or Tuesday. On Monday, we'll also have a few other posts, right? Yes. We have so many things and so many people wanted to share what they're building that we had to spill over and extend it to Monday. We might have something for Tuesday as well. It will be extended. It's not the first time. Sometimes that happens. Although for a while now, we have kept it from Monday to Friday. In the past, and birthday weeks are an example of that. We go over because there's so many things. And agents, developers, there's a lot there. So it makes perfect sense. Last week, we spoke about how the week is all about Cloudflare being the best place to build, deploy, secure agents. So we're actually bringing the full platform infrastructure not only to developers, but also to agents, and that helps the developers work, but also everyone working with agents these days. We're all developers, even non-developers in a sense. What can we say in terms of now that the blog posts are out? What is the main takeaways that people should take from this full week? Who wants to start? I can start. Yeah, I think the main takeaway, or at least kind of how I see it, is that as you're building agents, agents really need a different kind of infrastructure to run. They're fundamentally just very different kinds of software than kind of like the cloud, the 1.0 version of the cloud was designed to serve. And in this new world, you're going to need new primitives and just kind of new tooling to support this very, very new use case and this thing happening at such a larger scale. And I think the week was largely about this is what Cloudflare is providing for this new era of software. Makes sense. Any addition, Annie? Yeah, I think one of my favorite parts about the welcome post that Rita and Dane had put out was like they did like a quick math. What is the scale of where we expect, you know, agent infrastructure to need? And it's, you know, if we have X amount of knowledge workers around the U.S., each of them are running X amount of agents, then it comes down to around, you know, needing millions of compute instances to run these agents. And what's really interesting is like in the previous model, it kind of serves many users. But when you have an agent that needs its own compute instance to be able to run itself and also run the code that it's running, that scale becomes much greater. How do we do that sustainably? And I feel with the platform we've built on workers, it's unknowingly we kind of built the next generation of infrastructure. Exactly. Actually, I was hearing today from Kenton Varda on Twitter. He was mentioning how and he was the creator at start of workers nine years ago. And he was actually mentioning how he created containers, sandboxes, dynamic work, things like that, that at the time were not completely relevant for this age agents because this age wasn't here. But now those are actually really perfect for this. So it's interesting how things that were done a while ago are now actually perfect for this age. This reminds me, and in this podcast we had many shows, many episodes, including with John Wemm coming, which I started this podcast with, talking about how the Internet wasn't built initially for security, for a number of things that were added. Now we're in the agents era in a sense, and it was not all also built for agents initially, but with the layers, some already there, others are just improving, it could actually be better in terms of efficiency now because there will be more Internet traffic in general because agents will be performing more and more things. But also in many other things, it's really more relevant than ever, some tools that were around and new tools that are coming. So quite interesting to see the week forming that perspective in particular, and the blog post from Rita and Dane you mentioned defines that specifically. For sure. One of the things actually that the week starts, it's about Cloudflare's agentic cloud, right? Mostly about from developers to coding agents now and the importance of the agentic cloud. What can we say about that first day, that Monday, also with cool announcements? With our Monday announcements, I think the big story we wanted to tell around it was really focused on the compute primitive and all the different compute primitives you could use to build your agents with. We also slipped in something really interesting there around CLI for all of Cloudflare. So I guess the one that I would really want to highlight is our sandboxes, our containers. They went GA, and we actually have two announcements around that. So one is we have way more capabilities built into the sandbox SDK now where you can very much almost give your agent a full computer. So it needs a file system. It needs to run terminal commands. It can really easily do that with sandboxes. And I think what's really interesting, kind of going back to the story you told around the first generation of the Internet wasn't really built with a security mindset, I would say that we're seeing that with agents as well. Because the industry is moving so fast, people are starting to realize, oh, shoot, we really need to kind of lock this down, have a way to monitor all the security elements, like what can an agent have access to, and so on and so forth. We've added Outbound Worker as an egress proxy to our sandboxes so that you can get that monitoring for what your agents are trying to get access to outside of the sandbox. There's many cool blogs this week for sure, and this day for sure, and what those make people be able to do, agents be able to do for people in a sense, it's quite astonishing. For those who don't know, for example, how important it is building a CLI for all Cloudflare, for example. This has some repercussions online. What can you explain, folks, about it? Yeah, I saw so much excitement around people hearing that we'll not have a CLI for all of Cloudflare. So for those who didn't know, for a while I feel our CLI experience has been a little disjointed in the sense that for the developer platform we had Wrangler, and Wrangler only contained APIs for the developer platform, but Cloudflare is so much more. We also have application security performance products as well, and so with the Cloudflare CLI, this is a very early technical preview of what we're trying to build. We'll have CLI experience for every single product on Cloudflare, and so what makes this super exciting is that agents almost know how to run commands really easily, and so when you're able to give this one simple interface to interact with Cloudflare, agents will really easily be able to use it to spin up workers, set up a domain for it, be able to put CDN in front of it all with Cloudflare CLI. It's quite impressive. For those that are less technical, the thing that I find more interesting about this is, this is technical, but what this enables is not that technical. It's actually making things really easy and fast, and you having the ability, or the agent for you actually, having the ability to go over the products and the API operations, and there are so many there, with an ease that takes a lot from you, of you needing to understand and try to do specific things. It's quite interesting to see. It's all about also the single code mode MCP server, specifically here, and Cloudflare's entire API is really big, so this makes a difference, right? Yeah, exactly, and I think as we continue to work on it, we'll add more to it as well, so if you have any feedback for the early technical preview, let us know on Discord, let us know on socials, and we'll be sure to incorporate it. Should we move on to Tuesday? What can we say there? So Tuesday, very much focused on kind of now the security story. Like I kind of mentioned, especially if you're trying to use agents on an enterprise scale, you really need to understand what it has access to. So, for example, if you want to have an agent that writes code for you or employs resources for you, you might be giving it access to internal databases, infrastructure, you lock that down, and how do you make sure it's doing it in a way that you want it to? And so I think we have a couple of really interesting announcements here. The big call-out for the day, I would say, is Cloudflare Mesh. This one is, I think, the one that a lot of people might have heard already. It's our own private networking service that lets you connect your agents, your devices, all together in a private network. And so any sort of access between these two nodes would be happening within a private security environment. Those are also relevant for sure. In terms of the Cloudflare Mesh, it's also aggregating, right? In terms of workflows, in terms of abilities, right? You can connect various services to it, and it can all talk to each other within this mesh network. So you can connect services from private VPCs you have on AWS. You can connect it to services within the Cloudflare developer platform. You can connect your own personal devices, and then they can all talk to each other. I'm not sure if that answered your question. Sure, sure. But I've also seen a bit of repercussion there in terms of Cloudflare Mesh in action. People connecting a virtual machine, a laptop, and an agent's SDK project running on Cloudflare Workers to the same private network for secure access, for example. There's a bunch of examples that many folks from Cloudflare and others have been sharing on Twitter that is quite interesting to see as well. The ability of what is put together, in a sense, right? Yeah. The way that I think is really interesting to understand from maybe more of an individual, personal scale is a lot of people running OpenClaw a few months ago. In order to do that securely, have your OpenClaw be able to talk to, let's say, your mobile device. Cloudflare Mesh would be the product you would want to have to have that running in a secure environment. Exactly, and people can get started. Usually, all of these posts have a call to action at the end for people just to dip right in and try it out, in a sense. Anything we should mention more about Tuesday? I think another announcement I personally find super interesting is the one that we have around scaling MCP adoption. In this blog post, we share how we personally run MCPs in our own Cloudflare as a company. We have over 4,000 employees, and each of them may need access to the GitLab when they're using a coding agent, like Cloud Code or OpenCode. We share a reference architecture for how we set all of that up and how it gives us the ability to monitor and control what MCP servers our employees have access to and make sure unauthorized servers don't get used. That's interesting as well. Actually, this week I spoke with Sharon Goldberg because we had a World Quantum episode on Tuesday, because it was World Quantum Day. And I spoke with Sharon Goldberg and wrote this blog post about this specific topic in a very short way, but also a cool perspective. What should we highlight more? If you scroll up, let me see. I think there was also a really interesting one on Tuesday around manage OAuth access, make internal apps agent -ready in one click. That one was really interesting because if you're familiar with Cloudflare access, it allows you to enable a gateway in front of your applications that you want to be secured. You don't want anyone from the Internet to have access to. And with access, your users can enter a password and username in order to authenticate that they should really have access to this application. Well, now what's really interesting is it's not just humans accessing apps and the web anymore. It's also agents. And so when agents encounter an access wall such as the one we have at Cloudflare, typically they don't know what to do. They don't know your password. They don't have your identity. They can't get into this app. And so what will be highlighted here is essentially this new managed OAuth system. And it follows, I believe, a standard protocol, which will give agents the ability to also get access to these internal applications if you permit it to. And this blog outlines exactly how that happens and how you can give your agent the credentials and needs to access these apps. Yeah, it will save a lot of time doing that specifically. But you need to trust it, of course. What more can we say potentially? Maybe going over to Wednesday? Agents need all of these things in order to have the capacity to do real tasks. And that's what Wednesday and Thursday are all about. So I guess for Wednesday, looking at some of the highlights, I think one of the big announcements was Project Think, which is kind of Cloudflare's next evolution for our agents SDK. And I think there's a couple of really interesting parts to that. I think one is basically giving a framework for building long-running agents. Most things that people are doing with agents, like if you want to handle a real-world workflow, is not something that takes a couple milliseconds and then you're finished. It's things that can take minutes or potentially even hours. And so this is a way for you to build agents with really durable execution, where if it suddenly gets shut down, it can kind of pick up where it left off. The other, I think, really interesting benefit or angle for Project Think is this tiered execution environment concept. And that's all more about security. Basically, the ability to give agents the right tools with the appropriate amount of permissions to do the thing that they need to. So, for example, if you're doing something that only requires you to... If you're processing something local, you're munging some data, you might want to just run code, but you don't actually need to access the Internet. Whereas if you're doing something that requires the whole operating system and requires access to the Internet, you might want to give it a sandbox. And so having these different execution environments available means that there's always the appropriately secured component for the task. So, yeah, I think that's something that people were really excited about for Wednesday. And it has the safe sandboxes there helping keeping things protected, as you said, which is also cool. So, yeah, that one also had a great repercussion online. There's excitement there for sure. Yeah, yeah, definitely. I think like, yeah, some of the other cool things for Wednesday is our improvements to browser run. So I think previously named browser rendering, we're getting a new name browser run. But basically, I think agents also need access to a browser to be able to fully realize all their capabilities just because there are certain things you can only do in the browser. And what we've done with the improvements to browser run is add a lot more observability features. So as a person running the agent, we've added live view, so you can kind of see what the agent is doing in the browser and just make debugging like a lot more easier, recorded sessions, things like that. And then also there's now the ability for the agent to kind of hand off certain tasks to a human and then kind of then go back to its flow once the human's completed its step. So, for example, if the agent hits a captcha that needs human intervention, we now have the mechanisms to do that. We've also released within these new set of features like integrations with the Chrome DevTools protocol. So now different clients, agents like Claude and Cursor and stuff can kind of have access to these APIs and better control the browser through these things. And so, yeah, I think it's basically we've made browser run something that agents can use really effectively. We've also increased the limits on the number of concurrent browsers you could have. Kind of, again, taking back to this new scale that agents have or will need, rather, in this new world. Really interesting also this part of human-in-the-loop perspective where the human can step in and interact with the page directly to click or type or navigate. Even this week, I think it was OpenAI that launched a bit of a product that allows for its agent, in a sense, to be performing and moving around apps and people can interact with that as well. So it's definitely an area that will potentially put computers more to use where you can have your agents do something and you also do something as well, in a sense, right? Yeah, yeah. Maybe the last thing I just want to highlight is Agent Lee, which is Cloudflare's agent that lives within the dashboard. And I think this is exciting for a couple of reasons. One, because Agent Lee is an agent that we've built on top of all of our primitives or all of our building blocks that we're also offering to customers. So I think it's a cool way for people to see and kind of just get inspiration of what is possible when you're building on Cloudflare. And so for Agent Lee, we use a lot of the things that we've talked about previously in the blog. Code mode with MCPs. It uses a lot of the models that we offer through Workers AI. Yeah, we use our MCP servers. It's built on the Agents SDK and Durable Objects. And we share some of the architecture and how we've built it in the blog post for other people who are looking to build many similar types of agents. And then the other kind of really cool thing is it's just a really useful tool for customers of Cloudflare. I think particularly for people who are using a bunch of different Cloudflare services. Like if you have an app that you've got a worker for and you're using R2 and all these things. When something goes wrong or if you're trying to diagnose something or just find out information, you can use Agent Lee. And Agent Lee can use the MCP and use our tools to answer your questions rather than you having to click around in the dashboard yourself. That's super cool. The generative UI stuff is awesome to see. Basically building your own. It's kind of the beginning of building customized UI for users for your specific use case and task. So, yeah, I think Agent Lee is something that I'm excited for it to roll out to more accounts. Actually, there's some examples here with natural language in terms of how you can use it. And one of the interesting facts is that Agent Lee has that name because it's named after Lee Holloway, one of the founders of Cloudflare. There's a Wired story that people can find. Just search Lee Holloway and Wired and you'll see that story. It's quite a very incredible story for many reasons. So we can learn about Lee Holloway a bit and why we're actually naming it Agent Lee specifically with that story. Well, where should we go next? One, actually, close to heart, to be honest. It's the email service because I've started using it for a few weeks now internally for a project called Randomness Playground, all about entropy and walls of entropy and lava lamps. And it works really well. It works really well. Yeah, yeah. So we can move on to some of the stuff that's coming on Thursday. Yeah, email service. I think we saw so much chatter on Twitter and stuff like that about it. So many people, I think, were waiting for the open beta of our email service. So I think that was really exciting for a lot of the folks. I mean, it makes total sense. Email itself as a protocol has been around longer than the web has. And it's just kind of like this default way to communicate between different applications. Like a notification, you can kind of think of it as it is kind of like an API call email in a way, right? But structured email is kind of like an API call. And, yeah, now the email service is in public beta. I think the really interesting like agents angle to this is we've also, it has an integration with agents SDK. And agents SDK has an on email hook. So you can have basically an agent assigned to a email inbox, like have all your email sent to a certain alias go to an agent for processing. So, for example, if you set up a support at company.com email address and you can really easily then funnel all of those emails that go to there to your support agent to then take actions on. And so, yeah, that's, and I think we've also open sourced kind of like an email client that we've built to kind of show how this might work for people who are also looking to build agentic workflows that like are email based. And actually I saw some folks on Twitter saying, oh, I didn't know that Cloudflare actually had an email service before email routing. And it's one of those services that has many utilities, but this is now the full picture with email sending. You'll have the full picture and it's fully open and open source and free. Actually, you have a lot of you can do with a free version for it. Always great. And of course, the deploy to Cloudflare button for people to try it out. And to be honest, it's also close to hard because the team that built it is here in Lisbon. So I've seen them talk about it many times, which is cool. Where should we go next? Yeah, I think the other really, really big story on Thursday that people were really excited about was Artifacts, which is our new storage primitive, basically version storage that is like Git compatible. Yeah, I think people also saw a lot of excitement on social media and stuff like that for Artifacts. This is basically a new storage primitive that we've released for agents. And it's like a version file system that speaks Git. And it's really built for agent scale. So I think a lot of version control systems today are kind of built for humans using it. Humans who are maybe writing a couple of PRs a day, a dozen commits, whereas agents can potentially commit code thousands of times within an hour. And so that's what Artifacts is for. It's a way to store the code that an agent might generate, but not just code, like anything that the agent is generating where you might want version control, where you might want to be able to kind of like look at the different changes that the agent has made, be able to go back in time and start from like a previous version, be able to like give it to someone else and have them fork it. And so, yeah, Artifacts is all about like doing this at scale for agents. It's definitely ready and ready for what agents can do. Very well prepared there. There's a few examples here about even how much code that will be generated because of this new age. And that will have consequences and this will help deal with the possibilities here. So really cool one for sure. And on Hacker News, it was email service and also Artifacts and also another one, Cloudflare's AI platform and inference layer designed for agents. Those were really hot with people commenting and really hot on Hacker News. Yeah, the AI platform was one of the ones that my team and myself were working on it. You know, I think whenever something goes on Hacker News, I'm always like a little scared to read the comments because you never know what their take is going to be. So it's good to see some positive comments about that. But yeah, we're basically kind of like, you know, reframing AI gateway as this like unified inference layer. And so the big kind of announcements here are a huge expansion to our model catalog to a bunch of different model providers. People like OpenAI and then also a lot of image model providers like Minimax, ByteDance, also some like speech to text models. And so now, you know, it's not just LLMs that you can use through AI gateway, but like all sorts of different kinds of models. You can really build these like multimodal applications and you can get built for all of them through Cloudflare. And just, you know, it's the developer experience with the workers binding is just really, really simple. If you want to switch models, it's just like changing one line of code. So, yeah, this one is, you know, I'm obviously biased, but it's very near and dear to my heart. Ming, I saw you posted after that. Do we already have like a Cloud Opus 4.0? Yeah. Yeah. Yeah. So, yeah. Also, you know, because, you know, we've got all of these models. The model catalog is not static, obviously. And so, you know, our goal now is like to have the day that these new models drop and have it available to run on Cloudflare for our customers. And it's great. We have it internally as well. And it's always great to have the new models there. And I like this little feature, which is with AI Gateway, you'll get one centralized place to monitor and manage AI spend. Always important to see how much you're spending. And the ability to change models or bring your model is quite important, I would say, because there are models that are more expensive. You do a few tasks with them. You can do another task with other models that are less expensive. Also important in this day and age, I would say. And there's actually a Cloudflare TV segment all about this with Craig and Yuming that people can also check for sure. Yeah. If you don't want to read the blog if video is more your thing, we've got you covered. Exactly. Where should we go next? Yeah. I think those are the highlights for Thursday. So, we can go to Friday, which is the stuff that just came out this morning. I think one of the really exciting announcements for Friday is Flagship, which is Cloudflare's new product that Cloudflare is launching, which is our feature flagging system. I think the cool use case with Flagship is if you think about agents are writing so much code now, people are essentially just having agents build their entire application. And this really makes the loop of agents writing a code, them deploying it to production and testing it in the browser and seeing what works and looking at the logs. Makes it even easier to do or rather safer to do. So, you can imagine agent writing a bunch of code, then feature flagging just themselves into that new thing that they've just built. And then be able to then test it within a browser and click around. That can really speed up this development loop that the agent has, but all while not affecting your maybe main cohort of customers and still keeping them on the stable version. I think that's a really interesting use case for this feature flagging system. And one of the things I find really interesting is that many of the things that we spoke about before that are Cloudflare related, which is we're building many of these things with Cloudflare tools, are what is making some of these tools also really relevant. For example, in this case, we're using to not break things, things that Cloudflare does specifically. So, it's building Cloudflare with Cloudflare with a lot of advantages in terms of durable objects, Kiwi. So, there's a bunch of things of building Cloudflare with Cloudflare that actually makes perfect sense for products like these. I saw someone actually saying online that in this case, your AI agent can write your code in minutes, but someone still has to review, merge, deploy, monitor, and the agent can do that for you. And this method allows you to move fast. And if you want to see, you can see and review yourself, but not breaking stuff. Always important there as well, right? Yeah. And on your point of using Cloudflare to build Cloudflare, I think part of the reason why we're able to ship so many things that Cloudflare moves so quickly is because we have really, really strong primitives that we're building on, like durable objects and workers and stuff like that. And so, we're able to ship things like this new feature flagging service really quickly. But then also, the things that we build, a lot of it is born out of a need ourselves. Obviously, internally, we have a need to do feature flagging, and we also want to be able to run this nice little agent software development loop as well. And so, a lot of times, we're trying to solve our own problems. And Friday is all about empowering the agentic web, in a sense. There's also a few other announcements that are great. This is one that we publish every developer week and every innovation week, which is network performance update, trying to share how we improve the network. But there's other less frequent announcements. I really enjoy this one, which is introducing the agent readiness core. There's actually a website for that. You can actually put your website into this website, and it will give you how agent ready your website is. Which is itsagentready.com, right? I checked out this. I put my own personal website into the readiness checker, and I got 20. Not very good. But what I love about this tool here is it tells you exactly how to improve your score. I just included here celso.io. Celso is the engineering manager of the team that actually did this, so I put his blog. And his blog is actually pretty great. Yeah, people can definitely play around. And I really like this feature where you can improve your score and just copy all the instruction and put in your agent of choice. And it will improve your website giving the feedback, which is great. Yeah, very meta. Very meta, exactly. Anything we should mention about Friday as well? One story I personally was really fascinated by is shared dictionaries compression that keeps with the agentic web. So this one, I think it actually was while we were still working a week before agents week. That's when this announcement came in because it was a relatively new concept. I think, like, recently was enabled within, like, the Chrome browsers, but I could be wrong about that. What I think is really cool about this particular announcement is that, you know, especially as more agents are writing more of our web applications and, you know, applications, web apps in general are just getting bigger and bigger. We need a way to kind of, like, deliver these applications to the customer as efficiently as possible. And so with a compression dictionary, essentially we're comparing what kind of parts of the website that the client already has, already has cached. And we're only delivering, the next time you come to a website, we only need to kind of, like, give you the difference, even if it's, like, the website has updated. And so it makes your web apps be served, like, faster and also more, like, efficiently. And so I think this one will really be impactful in terms of, you know, the next era of the web where both agents are writing more code and also because agents will also be traversing the web and visiting websites as well. I don't resist, let me share this tweet from Elektrivik. They actually wrote a blog about the difference in terms of this feature, this product can do, like 99.5% smaller, compression just got a memory. So that's really important in this day and age, like savings, efficiency, all of that will be more relevant now that agents are doing more in terms of what they can achieve, right? Anything we want to mention? Another, actually, that had some attention was the introducing agent memory, managed service that gives AI agents persistent memory, allowing them to recall what matters, forget what doesn't, and get smarter. That's also a good perspective, right? Yeah, I can speak a little bit about agent memory. So I think it's super interesting because there's just so much discussion around how, you know, when you're building an agent, how do you manage its context, right? Some people call it, like, context engineering. And, like, there's been a lot of ways in which people are discussing, like, how memory should work within a harness. And so what we have here is, like, we have kind of, like, a fully managed, like, out-of-the -box product that you can directly leverage to give your agent everything it needs to manage its context with, like, an end user. So one component about it is, like, via retrieval and search. So, like, if you, you know, have conversations you previously talked about before, then it can kind of, like, give you that search through that previous context and fetch what it needs to use it within, you know, a given prompt or task. And then it also has, like, built-in, like, compaction, or some people call it, like, auto-dreaming, which helps you, like, as you're having more conversations with the user, like, the previous context is summarized so that it can be leveraged for the future without kind of cramming up the context window. I think agent memory will be really useful for anyone trying to build any sort of agents. These are the announcements so far. Anything we could mention about the Monday ones? The ones coming on Monday, I think some of them will be doing a little bit more of a deeper dive around how Cloudflare sells, how we use, how we've set up our own, like, AI and, like, agentic coding stack. I think that, you know, so many companies out there are trying to do this themselves. Everyone's trying to deploy agents. Everyone's trying to figure out how do we make our engineering team or our entire team generally more productive with agents. It takes a lot of, you know, I think everyone's kind of trying to figure this out, like, once, and, you know, what are the best practices, and I think, you know, we're just going to – we'll be sharing some of, like, what we do, the things that we've learned from trying to set this up, what are the best practices that we've found. So, I think, you know, if you're trying to do something like this within your own organization, that'll be really, really interesting and good to read. Absolutely, and to be honest, the way we do things is there's a lot of things that agents bring that people don't control and don't understand too much. Security elements in place, may that be – we talked about containers, sandbox, but even AI gateway, things like that. All of those things as an ecosystem will bring more protection, will bring actually at ease if you know that you can explore things in a safe environment and play around with it knowing that you're safe there if you do those type of things. And those examples, sharing those examples is always interesting for sure. So, let's stay tuned for that. When this is out, those blogs will already be out, so stay tuned for that for sure. In terms of the feedback, any surprises that you got from the week? We already mentioned some of the popular ones, but anything that surprised you in terms of even customers being really eager for some of the announcements? I really enjoyed watching, seeing everyone on Twitter and stuff try out, is my site agent ready and posting their own scores. And I think people were posting some pretty low scores, but then also some people who had a 100% agent readiness. And I think that just speaks to the kind of wide range that how ready people are for this new era. But yeah, I always get a kick of these things that are interactive and you can try it out on your own site. Yeah, this one isn't, I guess, related to a specific announcement, but I saw a couple of threads on Twitter where people are like, what is going on with Cloudflare this week? Like some people might not have been fully included and we have like agents innovation week. And so they're just seeing like a bunch of announcements come out and they're like, what's in the waters? What are they feeding the people at Cloudflare? So, I thought that was kind of funny as well. For sure, that score, I was really close because the team is here in Lisbon. Andres Jesus, actually, I usually work next to him. And I saw even on Twitter today, people sharing their bad results. And like 20 minutes after sharing, hey, I just gave this to my agent and now the score is like great. So how fast you can actually do improvements these days, it's quite astonishing for sure. And this is a good example. Observability in this case also, it's helpful for what you want to achieve specifically. So definitely a great one. I won't resist, but there's Cloudflare Radar also has some insights related to that announcement, adoption of AI agent standards. For example, how many are using robot.txt or sitemap AI rules in robot .txt. There's a bunch of cool trends here that people can monitor on Radar. That is pretty cool, to be honest. And URL scanner actually has also a feature related to those types of scans, which are great. People can explore as well. Oh, this has been great. Anything we want to add for folks to be attentive for what's coming? I guess those stay tuned for Monday and Tuesday because we still have a few things up in the pipeline. And even then, beyond, I think we're always cranking out new features, new products. And so even though this is the official end of agent week, I don't think we'll ever really stop continuing building and hopefully giving everyone what they need for their future agent workload. For sure. Ming, anything you want to add? Yeah, I think this is agents week, but I don't know, 2026 is probably like agents year, to be honest. And yeah, a lot of kind of like what, at least kind of I know for my team and I'm sure for a lot of other teams on Cloudflare, what we're thinking about for the rest of the whole year is like, how do we build better tools for people building agents? And so, yeah, I expect a lot more like agent related things coming out the rest of the year as well. For sure. To be honest, it's April and it seems like already a full year of announcements of things. And we're just like not in the middle of the year still. So what a year. That would be a good expression there. Thank you for doing it and stay tuned. Geek out. Thank you. And that's a wrap. So before we go, here's a video from my colleague, Zeke, from our developer team. So Zeke made this deep fake video. It's also a deep fake audio because it's mimicking Zeke's voice, but it's not Zeke's voice specifically. Zeke used many models for different things, video, audio generation, dialogue, sound effects, a bunch of things. It uses replicated models available on Replicate. Replicate is a company that Cloudflare acquired in late 2025. So it's definitely worth a watch. It only covers four days of Agents Week up until Thursday, not Friday. But it's actually a very cool sum of not only in terms of video, but also in terms of what he says. So here it is. Here's Zeke. It's Agents Week at Cloudflare. The week's not over yet, but there's already too much to keep up with. Zeke couldn't keep up, so he deployed me as his deep fake to give you the scoop. Let's cover the highlights. Here's the thing about Agents. Every traditional app serves many users from one server. Agents flip that. One user, one agent, one task. Scale that to millions of people and you need millions of simultaneous sessions. Containers can't do that. Not at a price anyone would pay. So Cloudflare built something different. Project Think is the next version of the Agents SDK. Your agent can crash and recover. It can hibernate when idle and cost you nothing. It can spawn sub-agents and escalate from a lightweight isolate to a full container when it needs one. 10,000 agents on containers means 10,000 always -on instances. On durable objects, maybe 100 are active at any moment. That's the difference. Sandboxes are now generally available to everyone. Your agent gets a real computer, terminal, code interpreter, live preview URLs, secure credential injection. Figma is already using them for Figma Make. Durable objects got facets. That means every AI-generated app can have its own SQLite database supervised by your code. If you're building a platform where users can vibe code their own apps, each app gets its own isolated state. Browser Run lets your agent control a headless browser with a live view. If the agent gets stuck on a login page, it hands off to a human. The human logs in. The agent picks back up. You get session recordings, direct CDP access, and support for up to 120 concurrent browser sessions. Voice agents got a new SDK. Wrap any agent class with WithVoice, and it can hear and speak in real-time over WebSocket. Built-in speech-to-text and text-to-speech providers, React hooks included. The Cloudflare email service entered public beta. Agents can send and receive email natively from workers. There's an on-email hook in the agent's SDK, so your agent can receive a message, do hours of background work, and reply when it's done. AI. Search shipped as a retrieval primitive. Create search instances on the fly, upload documents, and query with hybrid, semantic, and keyword search. One binding, built-in storage and vector index, no external services needed. The registrar API is now in beta. Agents can search for available domains, check pricing, and register them programmatically. Three API calls, a few seconds. It's already wired into the Cloudflare MCP server. Artifacts is a versioned file system that speaks Git. Create a repo per agent session, fork sessions, time travel through state, import from GitHub, clone with standard Git tools. It's built on durable objects, and it's heading to public beta next month. The AI platform unified 70-plus models from 12 providers behind one API and one build. The Replicate team is now fully merged into Cloudflare. You can bring your own model via Cog, and AI Gateway now buffers streaming responses, so if your agent crashes midstream, it can reconnect without repaying for the inference. Cloudflare built a custom inference engine in Rust and shipped pre -filled decode disaggregation, speculative decoding, and cross-GPU KV cache sharing. The result? Large language models that are fast enough for real-time agent loops running on Cloudflare's own GPU fleet. There's a new unified CLI. Run npx.cf to manage any Cloudflare product from your terminal. It's backed by a new TypeScript schema system that generates CLI commands, config, bindings, and docs from a single source, plus a local explorer that lets you inspect your local dev state for KV, R2, D1, durable objects, and workflows. Managed OAuth lets you flip a switch and make any internal app behind Cloudflare access agent-ready, no code changes. And Cloudflare Mesh wires up private networking between your devices, servers, agents, and workers, so everything can talk to each other securely without a VPN. That's 20 blog posts, and it's only Thursday. One more day to go. I'm Zeke Steepfake. You stay classy, developers, and good luck keeping... Microsoft Mechanics