# Rethinking Claude Code Best Practices — Lydia Hallie — session 2026-11-11T17:30:00.000Z → 2026-11-11T17:50:00.000Z _71 transcript lines · 39 slides_ ## Transcript Yeah, have you all tried Fable 5.2 yet? Yeah, anyone? How is it? Like it? Yeah, Opus 5.5 better, worse? Different use cases, right? Yeah. Cheaper for sure. So this should be on that screen. Right? Are you seeing anything, Jace? Let me see. So this has to be on my end. We need a cloud for IT. Anything yet, Jace? I I did what you said. Okay. Did it just die? Can I unplug it real quick? Yeah, please. Did I just die? I think it just died. Died. Oh, my gosh. Okay. I mean, I don't know. I should have died to that battery. Yeah. Now it... Now it's... It's just... I don't think. Should I just do my talk later today? We fix it. We don't. I'm starving. We have a little break in the next guy. Okay. Y'all, my laptop just died. So I... Nothing is happening. I don't know what to do. We, oh, I just restarted. Anyway, I feel like I've just spent twenty minutes just waiting here. I don't wanna take the next guy, but TLDR, we went from being super in the loop with our Cloud Code sessions to, you know, running these long running sessions that run for hours in the Cloud. With these models, like, Opus five five, Fable five two that are so good at verifying themselves, the the role as a human has really shifted. Let's see. I don't think this is gonna work. Oh, it's so weird. I know. Also, today, make sure you come by our anthropic booth. We're there all day. We have good merch. We also have good demos and just fun stuff. We have, like, a photo booth there as well, which is fun. And you'll see me there as well if you have any questions. But first, you gotta make this work a problem. Ignore. Why why did it restart? Windows Server experience a problem? What? It was, oh, Apple. Okay. Now I just gotta... I just wanna make sure I'm not... People often wonder what the story is behind what's the future of engineering in this space. And it turns out that fixing the computer and making it go. It's it's always a fun story. Like, the more advanced technology gets, the harder it is to basically make it do your bidding. And Conferences are conferences. AV is AV, and laptops are always going to laptop. I'm told that this might be the first time that an anthropic employee has not been prodding an agent, I'm sorry, to wind up getting a computer to do that bidding. That's what I'm saying. When you claw for IT. I just need a little clawed robot helping me. Let's see. Make your present notes. Exactly. I don't wanna step on your toes, but I can also keep running my mouth. Okay. Some people run Agents, I run my mouth. It... There's a production joke in there somewhere, but I'm too tired to find it. Things work in theory, but not in production. That's, why I call the staging environment theory. All black. Okay. Can you show us in the DSNs, please? Oh, wait. It looks like you're up. Maybe. I just have to see. I have my glasses with me. Remember, every code review is adversarial if you're a big enough jerk. Alright. That's right. Yes. Wait. No. No. No. No. Oh, that's not. That's not him. Sorry. It was... Yeah. Because I... And, of course, the number one reason to found a company is that as a cofounder, Maine becomes your feature branch. Good luck everyone else. Alright. Can I? I wanna see that on my screen real quick. Sorry. Because, Yeah. I don't know why it's not showing. Where should I go to? We're gonna I just don't. They want to. Close our mic so everyone can't hear us talking. Okay. Okay. Next to you. Oh, okay. Please. I'll see you later. Hopefully, I'll come back. You'll hear from me. Slight change. In programming here. I believe our next speaker is Michelle Chen. Correct me if I'm wrong on that. No one is yanking the, the cord there. If you can give us about three minutes, we will, take a small pause here. Besides, what are you all doing listening to this? You shouldn't be poking and prodding agents and seeing how many more you can wind up going? The tokens aren't gonna burn themselves. We need more bandwidth. Oh, suddenly, the use case for local Makes itself extraordinarily clear. Conference Wi Fi. There we go. I've done it on planes before, and what made that fun was suddenly my laptop became what sounded like the third engine on the jet. But it worked, and it wrote better code than I did. That's not a high bar, but it does mean I'm not going back in time to be the old timey, typing things out by hand like some sort of agrarian farmer. So, there's that. Alright. Next up, we have Michelle Chen. Talking about Cloudflare is a NeoLab. Personally, I believe that it's a database instead because I think everything's a database when you hold it wrong. But for some reason, I wasn't consulted. Anyway, Michelle, thank you very much. Please come on up. What does it mean for Cloudflare to be a new kind of AI Lab? To explore new models, edge AI, and the ambition to become the agent cloud, please welcome Michelle Chen, group product manager at Cloudflare. Hi, everyone. My name is Michelle Chen. I'm so excited to be here today at AI Engineer Code Summit chatting a little bit more about why Cloudflare became a NeoLab. So at Cloudflare, I am the group PM. Here I run on the product team, working on products like Workers AI, AI Search, AI Gateway as well. And so today we're gonna go through this little journey of how CloudFlare came to be. You might have heard that we launched and trained some models recently. I'm gonna tell you a little bit more about why, and what you can be doing with it. And so we're going to go through a little bit of a philosophical journey. When I was preparing for this talk, I wanted to use this quote, and then Anthropic launched a blog post last weekend and used my quote, but I'm still going to use it anyways. Goes in one second. The quote is from Kierkegaard, and it says, Life can only be understood backwards, but it must be lived forwards. And I think this is a really, really apt quote here for us because at Cloudflare, we're a sixteen year old company, and we've done so many things over the years, but somehow we got all... Some of the pieces right, and we ended up having to become a NeoLab. So we're gonna go through a little bit journey. The first thing we're gonna do is look backwards and talk through Why Cloudflare had to, became a new lab. Give me one second here. Perfect, we're gonna go backwards, and talk through why Cloudflare became a new lab. So, Cloudflare is a sixteen year old CDN company. We just celebrated our sixteenth birthday, at birthday week this year, in September. And we first started as a CDN. So as a CDN company, what you have to Do is you have to serve content as quick to your users as possible. You sit between the user's request and the origin, and you help, sit in the middle. And you could do things like cache and protect, do security and things like that. So that's kind of how Cloudflare started. But CloudFlare's infrastructure had to be adaptable in order to become the CDN. And so what we did was sixteen years ago in 2010, we launched with our own infrastructure. Our fast CDN had to be everywhere, and this looks really different than a traditional hyperscaler where compute is Centralized in US East 1 or US West 1. Cloudflare's infrastructure is a completely different paradigm. We have data centers all around the world as close to the user as possible, and this is really important in becoming a CDN. But now we have these computers basically running running as close to the user as possible. So we thought to ourselves, what else can we be doing with these computers that people would want to build with? And so came compute. We started launching these compute primitives as well. And so in 2017, we launched a product called Workers, which is Basically, like, serverless functions that you can run. It contrasted VMs at that time where you had to rent a whole machine and pay for the whole machine, either you're using it or not using it. And so with workers, you just execute some code and then the worker goes away. You don't have to pay for it. It takes milliseconds to start up, not seconds as well. And then we follow with other compute primitives like durable objects, which are more stateful. And then in 2025, we launched containers as well because there are some workloads that do need something like a VM, and we want all of those workloads to be able to run on Cloudflare. So now, you know, 2017, start of our developer platform, we have these compute primitives. People started using it. People started building more like static websites and things like that. So we thought to ourselves, how do we elevate this and take it to the next level? So we introduced storage as well. And so we have these storage primitives on Cloudflare's developer platform. We have kv, we have r2 dot object story, we have databases like d1 as well. And the idea here is that people should be able to build full stack application. All on Cloudflare. Stateful applications where people can go on and start building, full applications and web apps and platforms on Cloudflare's developer platform. So we have this infrastructure, we added compute primitives to it, we added storage primitives to it, and then came 2023. And if you guys remember what happened in 2023, chat gpt kind of blew up a little bit and it changed how we think about AI. Right? And I remember it being birthday week. And, three weeks before birthday week, a lot of us sat together in a room and we were like, what should we do now that ChatDPC is on the rise? What can Cloudflare play in here with our infrastructure and our primitives in order to be in this AI game? So in 2023, we launched a product called Workers AI, where we added GPUs to our network. And so again, this is really distributed, where GPUs are running as close to the user as possible. It's not centralized like a US East 1 hyperscaler, for example, but it's running as close to the edge as possible. And we had a hypothesis when we started workers.ai. Our hypothesis was that models were going to get smaller, so that network latency was gonna become more and more important. We launched with small models that could fit on our small hardware around the world that were super distributed, and we thought that people would have models get better and faster and smarter, and that network latency was gonna play a much more bigger impact here. Well, actually the opposite happened, and models got larger and larger, so. We went from something like 7,000,000,000 parameters, which was frontier scale at the time, to being the 3,000,000,000,000 parameter models that were getting launched. And I think this is not so shocking when you look back in time, because, you know, as we're on the way to AGI, if you're following scaling law, as you train on more data, as the models get larger, they're more generalizable. And so on the path to AGI, it makes sense that we're pushing the frontier of research here in order to get larger and larger and smarter models. So... This goes against our hypothesis, right, that the models were gonna get smaller. The models actually ended up being larger and the network latency mattered less and less. So what did we do? We ended up pivoting. So Cloudflare got really good at serving these giant models as well. So in March 2026, we launched Kimi, and then we followed up with GLM and DeepSeek and these large trillion parameter models hosted on our infrastructure. And so this is serverless, people hit an API call and get a token back. Actually took us a while in order to decide to do this, right? March 2026 is actually pretty late in the game to serving this model, but we were trying to make a conscious decision here on does this actually make sense for us to be serving it? It didn't really fit with our infrastructure story, but it fit very well with the developer platform. So in 2026, we had all these primitives in place. We had the infrastructure, which is unique to CloudFlare, running close to users. We had compute primitives where people were building agents on top. So workers and durable objects, right? As your agent needs to complete a task, it goes on and it spins and does a task and then sleeps when it doesn't. We have storage primitives, so your agents can actually interact with data and storage. And then we had models, these large models that were powering agents. And so all of this fit into the Cloudflare agent cloud story. And so we started telling this narrative that we wanna be the agent cloud. We want agents to be deployed on us, powered by models on us, using our compute and data primitives, all built on our infrastructure. And our network. But it never felt quite right. We felt like we were always missing a piece here, because the infrastructure pattern didn't necessarily fit in with what Workers' AI was designed for. And so I'm just gonna sit here and give you my second philosophical quote, which is, A brisk, considering that I'm a girl in Chinese and AI, but I'm gonna give you this Chinese quote anyways. And, it basically says, which means everything is ready. Here. All that's missing is the east wind. And one of the things about this is like you can't really predict where the wind is going to come from or how it's going to come, but you need to be very reactive when it happens. And so in September 2026, the east wind came. TypeSafe launched a model called JEV and introduced this idea of system one models. And it contrasted what we typically knew with LLMs, which were autoregressive and taking a long time to generate tokens, it scales quadratically, etcetera. A system one model is Completely different paradigm, where it actually just takes one forward pass through the inputs and generates an output that you can use to make decisions with it. And so I really think that this came down to perfect timing. I think people were starting to get fed fed up a little bit with LLMs because they were so costly, it cost so many tokens, it's super time consuming. And I don't know about you, but when I read a reasoning trace, sometimes my blood pressure rises when the LLM doesn't get the right, the things right. And so I think we were sending everything to LLMs and started getting With it. And so when this system one and dev paradigm started coming on board, we were like, this makes so much sense for us to adopt, right? We don't have to send every request to an LLM, we can send some of it to a system one model. And so the system one models actually end up taking work off of the giant LLMs, and you give it to the small, fast models. And the two are supposed to work in tandem. They're supposed to complement each other and get to the end goal faster. And so if you look, we are back at these small and fast models. Right? This was the original. Workers' AI hypothesis. And so Cloudflare was not wrong here, we actually were just early. With these small and fast models, the network latency became important again. We want the model to be responding fast, but also the network latency to be even faster. And so we decided to do something here. But, JEV was closed source. Cloudflare couldn't actually host the JEV models and take advantage of our infrastructure. So what did we decide to do? We built ## Slides ### 00:00:07 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROPIC [QR code] ### 00:00:39 # Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROPIC [QR code linking to https://codesummit.ai/](https://codesummit.ai/) ### 00:01:08 # Rethinking Claude Code Best Practices ## AI Engineer Code Summit Presented by: Lydia Hallie / Member of Technical Staff ANTHROPIC [https://codesummit.ai/ ### 00:01:40 # Rethinking Claude Code Best Practices ## AI Engineer Code Summit Presented by Google DeepMind Lydia Hallie / Member of Technical Staff ANTHROP\C [https://live.deep ### 00:02:08 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROPIC Presented by Google DeepMind [Logo: AI Engineer Code ### 00:02:39 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Presented by Google DeepMind Lydia Hallie / Member of Technical Staff ANTHROPIC [QR code] ### 00:03:06 # AI Engineer Code Summit PRESENTED BY Google DeepMind ## Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROPIC [QR code] ### 00:03:37 # AI Engineer Code Summit PRESENTED BY Google DeepMind ## Rethinking Claude Code Best Practices ### Lydia Hallie / Member of Technical Staff ANTHROPIC [https://www.anthropic.com/](https://www.anthropic.com/) ### 00:04:08 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROPIC Presented by Google DeepMind [QR code] ### 00:04:40 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROP\C Presented by Google DeepMind LIVE TRANSLATION [ ### 00:05:08 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Lydia Hallie / Member of Technical Staff ANTHROP\C [QR code for live translation] ### 00:05:41 # AI Engineer ## Code Summit ### Presented by Google DeepMind [QR code for live translation] ### 00:06:08 # AI Engineer Code Summit ## Rethinking Claude Code Best Practices Presented by Google DeepMind Lydia Hallie / Member of Technical Staff ANTHROPIC [https://events.google. ### 00:06:40 # Rethinking Claude Code Best Practices ## Lydia Hallie / Member of Technical Staff ANTHROPIC [QR code] ### 00:07:08 # Rethinking Claude Code Best Practices ## AI Engineer Code Summit Lydia Hallie / Member of Technical Staff ANTHROPIC [https://www.anthropic.com/](https ### 00:08:12 - Google DeepMind - AI Engineer Code Summit - AWS - Databricks - G2i - Sonar - Datadog - Judgment Labs - Alexa - ### 00:08:40 # AI Engineer Code Summit ### PRESENTED BY Google DeepMind [A grid of sponsor logos including parallel, Unblocked, AWS, Pinecone, SENTRY, Cisco, Render, OpenAI, ### 00:09:13 # AI Engineer Code Summit - AWS - Devin - Google DeepMind - G2i - Databricks - FACTORY - OpenAI - Cisco - Exa ### 00:09:40 # AI Engineer Code Summit - AWS - Devin - G2i - Databricks - Google - OpenAI - Cisco - Facto - Judgment Lab ### 00:10:12 # AI Engineer Code Summit - AWS - Devin - Databricks - Google DeepMind - Factory - GenAI - Judgment Labs - Cloudflare - Exa ### 00:10:40 # AI Engineer Code Summit ## cloudflare is a neolab Michelle Chen / Group Product Manager, Cloudflare [QR code linking to an unknown URL] ### 00:11:12 # Cloudflare is a **neolab** Michelle Chen Group PM, AI Platform - AI Engineer Code Summit 2026 [Cloudflare logo] ### 00:11:40 # AI Engineer Code Summit Presented by Google DeepMind > "Life can only be understood backwards, but it must be lived forwards." > – Søren Kierkegaard ### cloudflare is ### 00:12:12 # Backwards . ## cloudflare is a neolab Michelle Chen / Group Product Manager [https://cloudflare.com/neolab](https://cloudflare.com/neolab) ### 00:12:40 2010 # Cloudflare starts as a **CDN** - **no CDN**: user → every request crosses the world → origin - **Cloudflare**: user ### 00:13:13 # The fastest CDN has to be **everywhere** 2010 BACKWARDS cloudflare is a neolab Michelle Chen / Group Product Manager [https://live.aiengineer. ### 00:13:40 # Compute: your code, running in every location ## 2017-2025 - **2017**: **Workers** - serverless ### 00:14:11 ### 2017-2025 ## Compute: your code, running in every location - **2017** - **Workers** - serverless functions - **2020** - **Durable Objects** - stateful, one per name - **2025** - **Containers** - any image, next to a Worker BACKWARDS ### cloudflare is a neolab Michelle Chen / Group Product Manager [https://cloudflare.com/workers](https://cloudflare.com/workers) [Cloudflare logo] [QR code linking to cloudflare.com/workers] ### 00:14:40 # Storage: your data, next to the compute ## 2018-2022 ### 2018 - **KV** - global key- ### 00:15:13 ### 2018-2022 # Storage: your data, next to the compute ## 2018 - **KV** - global key-value ### 00:15:40 2023 # Models get **smaller** ### Cloudflare's hypothesis BACKWARDS ### cloudflare is a neolab Michelle Chen / Group Product Manager [https ### 00:16:13 # Anyway, the opposite happened Total parameters of open-weight models, as published **BACKWARDS** cloudflare is a neolab Michelle Chen / Group Product Manager CLOUDFLARE ### 00:16:36 # Anyway, the opposite happened 2026 - **Llama 2 7B** (2023): 7B - **Llama 3.1** ### 00:17:12 # So Cloudflare got really good at serving giant models - Kimi - GLM - DeepSeek BACKWARDS cloudflare is a neolab --- **Presenter:** Michelle Chen ### 00:17:36 ## 2026 # By 2026, everything was ready for agents - **infrastructure** - network + GPUs - **compute** - Workers ### 00:18:12 2026 # By 2026, everything was ready for agents ### The Agent Cloud - **Infrastructure** - network + GPUs - **Compute** ### 00:18:35 > 万事俱备，只欠东风。 > “Everything is ready. All that’s missing is the **east wind**.” > > – Chinese proverb cloudflare is a neolab Michelle Chen / Group Product Manager [Cloudflare logo] [QR code linking to https://cloudflare.com/neolab] ### 00:19:08 2026 # Perfect timing: LLMs were **overkill** - Is it urgent? (decision) - Which team owns it? (decision) - What's the ### 00:19:36 # Perfect timing: LLMs were **overkill** - Is it urgent? (decision) - Which team owns it? (decision) - What's the total? (value)