Loop Engineering, OpenAI Sites & the Great China Model Shift | This Weeek In AI

This is a tweet from OpenAI, it says, Building apps has never been easier. With sites, codecs can turn your work ideas and plans into an interactive website or app. So this is like a Lovable or a Replit. The model used most on Lovable is the one that is going to destroy Lovable. Yeah, which is also wild because Lovable probably gave it all the training data. Exactly. To make an even better model for that task. Oh man, it's just the way it goes in this game. The way it goes with the Model Labs. Hello everyone and welcome to Agents Hour. We are in person this week, so that's a bonus. What's up dude? Nothing much. We're in the Code Rabbit office once again. I'm a lot better since we've been here. Feels cozy. Cozier here. Let's jump into the news today. Back at it again! There's always some drama. There's always things for us to talk about. Yep. First things first, 8x faster. Bots are greater than humans. Okay, what does that mean? Well, Anthropic released a report or at least a tweet that said, Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021 to 2025. So essentially there's a timeline here of releases and the argument that they're making based on, I would assume metrics, is per engineer, the average engineer is now shipping 8x as many lines of code. So we can argue if lines of code is a good metric, especially with AI able to generate it. But 8x as many lines of code per engineer getting shipped. I had a really cheeky comment to that. Let me see if I can share it. It was funny because someone internally said, you know, saw those HTML files and then you responded with this. Yeah, I was like, 55 HTML files, 55 React components, 55 MCB servers, 55 tool calls, 55 agents, 55 workflows, 55 emails, 55 prompts, 55 vector DBs, 55 cloud.md files, please. If you guys like Tim Robinson, you should check it out. But I just thought it was so perfect for this thing. It was, yeah. That's one way to ship 8x as much code. And you could argue that just because it's 8x as much doesn't mean it's actually 8x as productive. But I do agree with the sentiment though that on average, engineers are able to ship more with these models. I think that is true. Whether it's 8x as much, that seems a little... But we don't have, you know, maybe we don't have the Mythos preview access that they have yet. Yeah. Or, you know, 8x the code, but is it 8x the value? That's what I question. I would argue probably not. And then this post came out right after that. Anthropic employees are depressed. And this is probably anecdotal, but it got a ton of attention after the post that was kind of talking about how fast Anthropic is moving in general. And they're saying, well, their employees aren't very happy. And I don't know if I agree that on par they're not happy, but it's just kind of a funny anecdote that just because you're moving faster doesn't mean you're happier. Yeah. If you're watching live and you saw us talking to Juan from CodeRabbit, we talked about this a little bit, but agentic traffic is passing human traffic in a lot of ways on the internet, in docs, using CLIs. Matthew Prince from Cloudflare said, it happened faster than I predicted. Agentic traffic has now, so bots have now surpassed human traffic online for the first time in the internet's history. It's wild. First, he thought end of 2027, then early 2027, and now it's right middle of 2026. That's a fast on-ramp. Yeah. We got to talk about loop engineering. And this first one's funny because it was taken down. So there was a post about Boris, a video, Boris talking about looping. And I can't show you the post because it's been removed from X because apparently it wasn't supposed to be able to post it. But you watch the post. Tell the folks that are watching what was the post about. What was the sentiment around this idea of what loop engineering is? Yeah. So apparently in this video, there was an interview with Boris from Cloud Code. And he pretty much was like loop maxing on everybody, saying that he doesn't really prompt anymore. He tries to figure out the right loops to run. And now he's in this like loop engineering kind of phase, which I thought it was interesting. When you have unlimited tokens, you could do whatever you want. So that's totally cool. But it caused a lot of, once again, caused a lot of controversy, lots of talk on Twitter. And I'm thinking that Anthropic is not really down for all this Twitter hate. Maybe that's why they took down the post or something like that. But I'm really curious why they took the post down. Peter Seinberger also weighed in and said, here's your monthly reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents. It's a similar sentiment. And it happened within the same, you know, roughly couple days. And internally, I know we have some folks that are starting to experiment with this quite a bit more as well. We had Ruben in the chat ask, do you guys still prompt? Yes and no. I think we are designing loops ourselves. If you look at the last month of the coding agent wars, Anthropic came out with the slash workflow, which, you know, takes your input and turns it into a dynamic workflow, which can then be repeatable or not, if that's up to you, which is essentially a loop creation mechanism. You go from idea to deterministic workflow that can have a lot of parallelism up to 16 or so. I think we'll have a diagram that we'll show in a second. Yeah, maybe we should jump right into that. Yeah, let's just jump into it. We want to talk through kind of some of the evolution of what we'll call looping. Looping is what I'm talking about. Always be looping. Always be looping. So the traditional loop, I don't think this is dead. I use this still quite a bit. Yeah. You just send a prompt, maybe optionally you put it into a plan, maybe sometimes you don't. So the plan step is optional. Your code agent does some stuff. You provide some feedback. You look at it and say like, no, don't do that. Yes, do this. Or wait, we also need to add this. And then you loop. And then eventually you PR it. Maybe you have code rabbit review it or you review it or whatever. And then, you know, you repeat the cycle. That's the traditional loop. And this is what they're telling you not to do. Yes. Which I think is bullshit. Yeah. Like there are times I think more complex loops are valuable. Yeah. But to say you shouldn't be prompting anymore is just clickbaity. And I think it is, the answer is it depends on what you're doing. Yeah. If you're doing a bug fix that you know the answer to, just send it in this loop and you probably don't even have to provide a lot of feedback. Yeah, it might be like one turn. Yeah, you don't need a big goal or huge written plan. Yeah. And if it's something that's very important and you want to deeply understand it, you probably should ride the loop. So riding the loop means you're on the loop. You're providing the feedback. I think there are some times where you still want to be on the loop. Only the Sith speak in absolutes, right? Then if for those of you that have been following the space for a while, you know, there's this concept of the Ralph loop, right? Where you prompt, you create a plan, you decompose that plan into a set of tasks. Usually it's in a file. It could be in a, you know, some kind of DB or project management system, whatever. And then your code agent will continue to loop through the task list, completing off tasks, marking that it's completed the task. And it'll continue to loop with fresh context each time. And the idea around that was compaction was pretty bad. And you didn't want the code agent to maintain its, you know, state. So you wanted fresh context each time. But eventually the code agent would decide it completed all the tasks and the loop would finish. And sometimes you'd add like there's a progress file you could add. I mean, there's some different ways you could do it. But this is the general flow. Like your code agent is looping on itself and completing a task list. We used to do a lot of Ralph loops that none of the code actually emerged. But we were running hella loops for a while. It was good for prototypes. Yeah, it was the beginning of the year. Yeah, that was like December. That was like Christmas and New Year's timeframe. A lot of Ralph loops. Always be looping in that phase. Then goal came out. Yep. You know, slash goal is you prompt, you probably create a plan, but you don't have to. Yep. I usually recommend it. The more detailed the goal and the plan, the better. So you create a plan, you run slash goal, your code agent will try to complete the goal. Then you have some kind of judge or evaluation agent that checks, did you finish the plan? It'll critique it almost like it's reviewing it as it goes. And then eventually you go through the loop is complete and maybe you PR it. So that's a goal loop. And I use this quite a bit, especially for tasks that you want to run a little bit longer that you have more, you know, you can better define what you're looking for. You can kind of let it run and it might run for half an hour, an hour, and it'll do its thing. And this is where I think designing your loop comes really handy because goal loops can run for hours and hours and hours. I mean, I wrote, I ran a goal loop over the weekend, 12 hours, which is nothing compared to some people's goal loops of like 36 hours or whatever. But if we're anticipating a long running agent future, you know, it's going to require a very detailed plan, a very detailed goal, and then a really good judge. And then recently with Cloud Code coming out with dynamic workflows, let's talk through kind of what that looks like. You still do the, you prompt into a plan, you create some kind of plan. Most likely you have a code agent that then essentially creates a script. Yep. And there's all these like dashed lines because this could change every time because it's creating its own workflow a little bit more deterministically. It's spawning multiple parallel sub agents to accomplish that. It typically will have some kind of judge or like fixing agent that'll also be there at the end. And it'll try to basically write a workflow that will complete that task, whatever the task is. And it'll define a workflow based on what it thinks is best for completing that task. And Anthropic released like a bunch of different workflow structures that they're kind of setting up through these loops, right? This is just another form of a loop. Benefit here though, is these can be repeatable. So if you are doing long running tasks that are repetitive in some way, along the way you can assemble workflows that maybe, you know, just the input changes, but the whole structure is still the same. So I think there's some merit here in dynamic workflows as well. If you're designing your own loop, you might be designing workflows that depending on the types of tasks you may want to use for to accomplish those tasks. But I guess my question is, do you think there's need for people to go beyond what you can get with goal or dynamic workflows and craft it themselves? It sounds like it made it seem like Boris and, you know, even others on our team are kind of really meticulously planning out how they want this loop to work. Yeah. Do you think that's valuable for people or should they live in the just slash goal world right now? Our team is living in slash goal. I mean, we're working on the dynamic workflow piece. I have not seen the value of the dynamic workflow piece having have a prototype of it, but goal I've seen so much value. And maybe it's because we need to start daily driving dynamic workflows to see the value. We want to support both. Some tasks are fit for deterministic workflows. Some are really exploratory. You don't really know what the hell to do. So goal is really nice for that. And one Lopez says, what's next scheduled task engineering where you craft cron tasks to check on your loops? Hey, it might be a heartbeat. It might be a thing. Loop engineering is just a different name for AI engineering. It's just the new hype word. Yeah. It's getting better at your job. Pretty much. Yeah. It's like now there's some new tools. Maybe you can do things differently. Not it's not all one size fits all. Yeah. New patterns. Dude, did you subscribe? Dude, I host the show. Did you subscribe? Did you subscribe? Subscribe to Agents Hour every Monday, noon Pacific. So OpenAI, June 2nd. This is a tweet from OpenAI. It says, building apps has never been easier with sites. Codex can turn your work ideas and plans into an interactive website or app your team can explore, use and share with a URL. Rolling out to business and enterprise plans before expanding more broadly. So this is like a lovable or a replet for business and enterprise users. And then eventually probably for all users. I think we had a bunch of internal discussion about this, how pretty much lovable is cooked. Codex is going to take their market share. And I really liked this one quote, which was like, the model used most on lovable is the one that is going to destroy lovable. Yeah, which is also wild because lovable probably gave it all the training data. Exactly. To make an even better model for that task. Oh man. Which is just... Just the way it goes in this game. The way it goes with the model labs. They're trying to eat everyone's lunch. Cognition came out and said, AI should earn its keep. Introducing AI productivity guarantee. If Devin delivers less engineering value than you're paying for, Cognition will fund your usage until it does, up to 10 million. It's time for the AI industry to stop maximizing tokens and start maximizing productive output. So they want, seems like mostly for like enterprise users or whatever, they're basically willing to say, hey, we'll give you free tokens if it doesn't meet a value threshold. And I don't, I didn't read into exactly how to define it, but apparently they have some mechanism for defining what is productive output, which I think is kind of hard to define. Yeah. Maybe it's just merged PRs or accepted changes from Devin. This is pretty cool because there's a lot of conversation right now about people not getting value from the token spent. And I mean, we use Devin and there's definitely tons of tokens that are worthless that we've spent. Things that never merged. PRs that are open that are never merged. Never merged. So like that would, I would consider that not having value, right? Maybe we should get our money back on those. I mean, and I think there is value sometimes in unmerged code if you're prototyping. Yeah. But if the intention when you opened it was to actually make the fix and you don't make the fix, that is wasted. Yeah. Completely wasted. Your intention was to prototype something and learn from it, then who cares if it wasn't merged? That was your goal. Yeah. But I think there are many instances where we have open PRs and then they just sit there and rot because it's easy to open PRs now. Yeah. It's hard to get them merged through. Unless you use CodeRabbit or something. Introducing Devin Desktop. This is another one from Cognition. Managing fleets of local and cloud agents from one surface. Plan, delegate, review, and ship without leaving your editor. I think this is kind of what they started to pull like windsurf into, right? Yeah. WindSurf is dead. Yeah. But basically, I think they've tried to basically kill windsurf and- Now they have this. They have this. It's like their new thing. Like don't use windsurf. Use Devin Desktop instead. Dude, pour one out for windsurf. We used to use windsurf. I was a windsurfer. I was a windsurfer too. For a few months. RIP. For a couple months at least. Was it two months maybe? Then we turned into cursor dudes. Well, it was cursor, then it was windsurf, then it was back to cursor. Yeah. Pour one out. Pour one out for windsurf, yo. Cloudflare acquires VoidZero. So VoidZero is joining Cloudflare. Our mission stays the same to make JavaScript developers more productive than ever. So this is just interesting. Cloudflare making acquisitions, making moves. A lot of stuff happening with the Chinese models. I saw some other posts similar to this. It's a shift towards Chinese models by American AI startups. So this is from open router API calls. And you can see that, you know, since really January, it's really started to overtake U.S. models. So more Chinese model tokens than U.S. model tokens are being spent on open router. I believe it. His supporting evidence flow from Lindy said they pulled the trigger today and switched 100% of Lindy traffic to DeepSeek V4, churning from anthropic models. It saves us millions of dollars. And we're actually seeing an increase in performance on many use cases. Transformative for the business. Because DeepSeek is probably a tenth, a twentieth of the cost. Yeah. And if they're able to say, even if it's 95% on some things and maybe even more than 100% improvement on other things, it's probably worth it. Yeah, you don't need Opus for everything. These really expensive frontier models are going to be like the limiting factor for many businesses, you know. We are very much playing with open models as well. DeepSeek being one of our favorites, Quen 3.7 Max is also a very good one. We've been testing MiniMax. We've been, people on our team are using GLM. Yep. I mean, it's, we've been experimenting with all of them. And there are definitely cases where it's not as good, but there are oftentimes you don't need it to be as good. Yeah. You just need it to be capable. Yeah. And not slow. Notion came out and said they had to swap over from anthropic models because of anthropic, more anthropic downtime. Seems like even the new compute still has not completely mitigated their reliability issues over there. Maybe it's because they ship so much code. Yeah. Maybe they're running too many loops. There are too many loops. It's interesting because I imagine Notion will just turn it off temporarily, we'll go back when it's up. But you do have to think that if big companies like Notion are continuing to have issues with reliability, they're just going to start using these other models. The pull of Lindy. Bunch more models released. So we'll run through these pretty quick. If you're an Android user, you'll have AI on your phone. That's where all this is leading, right? These smaller models. Maybe that's the key for Android to maybe assert dominance. Yeah. In the short term. I feel like Android probably still, yeah. More people use Android than iPhone today or not? I don't know. I mean, globally. Globally, but maybe not in the US. But if it had better native AI support, maybe that's a way for people to consider an Android phone. Plus, if you're using a smaller parameter model on your phone and it actually is very capable, that just shows you that you do not need a Frontier model as well. Jemma introduced Magenta Real Time 2, an open model musicians can play as an instrument. So you can basically play it as an instrument, which is pretty cool if you're a musician. There was also a release last week called Miso 1. It's the most emotive voice model in the world. So people are still trying to innovate on voice models. It's pretty impressive what's already out there and the fact that there's even more models that are getting released. It's a good thing. Nematron 3.5 ASR streaming. NVIDIA is releasing a whole bunch just on the Nematron family. So this came out last week as well. Liquid AI introduced LFM 2.5. I'm not going to read that whole thing, but vision language models that return structured JSON, not freeform text. So it's like JSON specific structured responses. So you pass an image and a list of fields, get back a clean JSON object. I've been doing a lot of like macOS automation lately. And I think VLMs are going to be a huge part of the story. Who wants to take screenshots of shit, right? Like you should just know what's on the page. And I don't know the mechanisms for that yet. But you know, like Juan was saying that it's an ensemble of models put together. That's why it's important for all these different individual models to get better. You got to have really good voice. You got to have really good VLM. You got to have good, you know, PDF parsing. You got to have all these things need to be, need to come together. Yeah. From base 10, GLM 5.1 now achieves 160 tokens per second and a less than two second time to first token on base 10. So they're releasing some stuff over there to try to make inference faster. They're trying to compete directly with fireworks. Minimax M3 got faster. So not only new models, but people are trying to make these models fast, which makes sense. I know we had been testing Minimax M3 and your first comment was it's pretty good, but it's slow. Yeah. And then a couple of like a couple of days later, they said, we just made it faster. And it is not that much faster, but it's faster for sure. Yeah. They don't have the same hardware as we do in here in the mighty USA, but just use it on fireworks and you'll be fine. And then they, speaking of fireworks, Minimax M3 arrives with Minimax sparse attention. So it's 15.6 X faster at decoding 1 million tokens. And then they're partnering with it to power inference. So if you do want to try it out and use it on fireworks. All right. It's some funding. Suno raised a whole bunch of money. Good. Good. We need some new intro music and outro music. So make it even better. We'll use it. I feel like they just raised too. I know. It seems like it's been pretty recent. And then Superbase raised a bunch of money and Superbase's comment was, you know, basically most databases are created with agents by agents or used for agents. Another thing that's kind of on this topic is SpaceX is apparently going public on Friday. Friday. So that's P C X. Who's going to buy it? So if you're going to buy, you know, Anthropic was rumored to start the process, but SpaceX and, you know, plus XAI is part of that, right? It's all going public on Friday. And if you can't tell, I'll be dealing with a little cold. He's, you know, powering through here for the show. Airbnb CEO Brian Chesky started a new AI lab. So he's still going to remain the CEO of Airbnb, but he's basically starting a lab on the side, which is interesting. Yeah. Focus on design and UI. I mean, he's been talking about this even when we were in YC when he came to talk to us that he's really focused. Like he doesn't think AI is good at front end engineering or UI or anything. Yeah. And I think it's good at front end engineering and UI if you want everything to look the same. Exactly. So in that case, it's not very good at front end engineering. It'd be interesting to see some new UX patterns come out of whatever this is. But Tyler on our team is betting anyone $100 that if they do produce a model and it's better than open models today, he'll pay you a hundred bucks. Who? This? Anybody. Anybody? Yeah. Well, if Chesky could do it, which he does not believe they can. So. All right. I thought this was funny. There's this is AI profitable.com, which basically says no, AI is not profitable yet. Because if you are following the space, you know that oftentimes it's one AI company paying another AI company. It's just money moving around. Yeah. And so this tries to remove that from the equation. Say, if you didn't count an AI company paying another AI company, is there actually money? And the answer is today, no. There's money for one company in particular though. Yeah. That one company. Can you all guess which that one company is? It's all flowing to one. Mr. Wang's company. Let's go through the quick hits. So V0, you can now launch production ready Shopify storefronts without leaving V0. That's cool. Factory introduced model routing to factory. So now factory will pick the right model for the tasks. Again, kind of what we're talking about though, you don't always need the best model. Yeah. So it's going to do it for you. I mean, I think the negative to this is it certainly breaks the prompt cache, but if it's a really long running task and very expensive, it's probably okay. For daily tasks, maybe it's less, you know, just the prompt cache saves you some money. I feel like the prompt caching is becoming a crutch. Yeah. Unfortunately. No, I think everyone leans into like, oh, it's a prompt cache. It's because they discounted so much. Yeah. But if you could go from Opus to DeepSeek for some things, even cached Opus is more expensive than DeepSeek. Exactly. On par. So what are you supposed to do? Just like stick with it just because you have a prompt cache or do you do what the right thing to do is? I'm sure this whole caching thing will get solved too. Hermes agent introduced Hermes desktop. So everything you love about Hermes now native on your machine. So you can use it locally and not just, I guess you have your own desktop app now to run Hermes. And that's it. That's the show. It was pretty chill news week. You can of course follow us. You should go give us a star on GitHub. I think we're almost at 25,000 stars. Almost. Maybe this week. So maybe right now, if you're listening, go, you know, slam the star button if you haven't already. If you're somehow or watch have watched to the end of the show and you have not given Mastra. What are you doing? We go a star. I'd be surprised, but go give us a star. Follow us on YouTube. Subscribe. Follow us on X at Mastra. We have more stars than AISCK. Just throwing that out there. Yeah. Not by much though, but we need it guys. That's why we need your help. Come on. Let's go. You can follow me on XSMThomas3. You can follow Abhi at Abhi Iyer. This has been Agents Hour. We do this thing every week. We bring on guests. So if you watch the show, you notice we had two guests live because we're in the Code Rabbit office today, which was awesome. Sometimes we bring on virtual guests and we also talk about the news. I just want to give thanks to Code Rabbit. We all don't use Code Rabbit. We highly recommend it. We use it every day. It's kind of saved our asses for two big product releases. And like Shane said earlier on the show, we do not do code review until the rabbit is happy. It's a really good thing that's worked out for us. Hopefully it works out for you. Check out their change stack. Check out all the stuff they're doing. They have crazy customers, Indeed, Mostra, and many of the likes. So thanks Code Rabbit. And we'll hope to be back in the studio soon. All right, everyone. Thanks for tuning in to Agents Hour. We'll see you next time. Peace.