seeco - Head of Claude Code: "I'm not prompting my agents anymore I'm build...

This is sort of the crazy thing about building on the best model, is it advances extremely quickly. And so it got to this point where the model was able to write all the code very quickly. And that was sort of a leap that we made. And now the second leap that we're making kind of as of three months ago, and we're in the middle of right now, is to the point where agents are prompting agents. I'm not talking to agents, my agents are talking to other agents. And so you have these kind of big organizations of agents that are doing that. It's nice to be here and appreciate the invite. We wanted our investors, Boris, to really hear about software engineering directly from not only a software engineer, but a software engineer who actually created ClockCode. So it's really a unique opportunity. So thank you. So I'll just kind of dig into it quickly because I know our time is limited. But Boris, will you just give us the quick background on when you started Anthropic and how you developed ClockCode and a bit of just the history of how this groundbreaking product that everyone's kind of heard about was developed? Yeah, absolutely. So it was the mix of maybe like two different things. So one was at Anthropic, the focus has always been singular. It's always been on enterprise and coding in particular. And that's been the case for many years. So since we kind of figured out how to make the models work and figured out how to get them stepping and training and so on, that's always been the focus. When I joined Anthropic, we wanted to figure out how to productionize it and how to bring something to market that would actually show the soft. Because when I joined, there was this just like tremendous feeling of product overhang that the models can do these things that just no one in product has yet realized because no one is being ambitious enough with how far you push it. And so the team that I joined was called Anthropic Labs, and it was a prototyping team. It's like a little innovation product incubator within Anthropic. And we produced four things. So I built Cloud Code, another engineer built MCP, a third one built the desktop app, and a fourth one built skills. So you can kind of see the seeds of everything in Anthropic's approach to coding and approach to product. It all just started with this one team because the direction was just very obvious, even if the form factor was not obvious. And that took a lot of experimentation to get right. And yeah, that's kind of how it started. There was immediate product market fit internally. Externally, it was also a hit. Yeah, I think what people may miss generally about models and coding is how intricately intertwined they are together, right? It's almost like if you're in the construction business, but you're also in the scaffolding business, right? And obviously, scaffolding can directly impact your construction business and vice versa. And I think to your point, Boris, I think part of the reason why both businesses are doing so well is if we think about the future, right? Coding is a key part of how a model is built and is also a key part of the output that it produces. And so you kind of get synergies between both. And I think that's kind of the point that you're alluding to. Yeah, and honestly, just from a mission point of view, our mission is to build safe AGI. And the model has to interact with the world somehow. The model exists as software. And so the way it interacts with the world is through code. And so this is just it's such an important thing to get right. Maybe I think what'd be interesting is, you know, a lot of our investors are not software engineers, so they may not realize just how the work of a software engineer has transformed, right? And one of the things that struck me the last time we spoke is how much your role and job has changed, you know, as a software engineer over the past kind of even just kind of 18 months. Can you walk us through a little bit from when, okay, I used to write the code myself to then maybe I had autocomplete to maybe even give us a real-time view of how you do your job today? Yeah, when I first started coding, it was, you know, it was all by hand. And, you know, like my grandfather actually used to program with punch cards. So this is like before programs were even encoded in software, they were encoded these little pieces of paper. We actually just donated the first quad code, the first video demo of the quad code that we created to the Computer History Museum. And we kind of got a tour and kind of got to see the history of computing. And it's just amazing. Like, this is how my grandfather used to code. It was, you know, on paper. And then at some point in transition to software, and this is how I grew up. And you started to encode programs as source code. And that was the case for about 70 years. And, you know, like if you look back like a year ago, the way that people were coding was like autocomplete. It's like, you know, the same thing, like you're typing on your phone and it completes a word. That was the state of art for AI programming. And what happened was quad code introduced this idea of agent decoding. And what this means is the agent fully writes all the code. And from a user experience point of view, this was a step up. Because I'm describing to the agent what I want, and it writes the entire program for me, as opposed to me writing source code directly, like writing the document directly. But I think from like a long view, this is as big of a shift as the shift from punch cards to source code. This is actually like a once every 70 years kind of shift. And we actually might be about to... Are you even writing any more code anymore? Do you even type in the ID anymore? Or are you just kind of managing these agents now? I don't have an ID. Yeah, I haven't used an ID in about six months. Just for the audience, ID is essentially like the Microsoft Word for a writer. It's where you would go and kind of do things. And the reason that's so transformative now is basically you're saying, hey, I'm just managing these fleets of agents that are doing tasks for me. That's right. And these days, I think at any given moment, I have at minimum a few agents running. Often overnight, I'll run a few thousand agents to do some kind of complex work. This just was not feasible even three months ago. So this is where everything is headed. And so do you now... How do you manage these thousands of agents? Do you have agents that are now actually managing the other agents? It seems like, my gosh, if you have so many going, how do you even organize yourself? Yeah, this is sort of the crazy thing about building on the best model, is it advances extremely quickly. And so it got to this point where the model was able to write all the code very quickly. And that was sort of a leap that we made. And now the second leap that we're making kind of as of three months ago, and we're in the middle of right now, is to the point where agents are prompting agents. I am not talking to agents. My agents are talking to other agents. And so you have these kind of big organizations of agents that are doing things. And honestly, it happened very organically. Because what we were seeing is engineers are just running many, many agents. And they're the ones that demand some way to manage these agents more organically. So they don't have to do it. Yeah. So I think the punchline takeaway for the audience is that the revenue ramp that you're seeing, and I could understand as an outsider, the numbers are so big, and everyone says transformative and this and that. And maybe you kind of get drowned out a bit in the noise. But really, if you look at the practitioner level, and Boris is one of the top practitioners in the world, what we want to kind of communicate is how fundamentally the work is changing. And how it is switching from a human doing manual labor of putting in code in a window, to actually deploying fleets of agents that are now task-oriented and even kind of manage other agents. Maybe then my last question, Boris, as we move forward is, what do you think this means for the development of models and software generally? Are we going to have apps that are basically autonomous, that are fixing themselves, that are continuously developing new products on their own without even a user or a quote, owner prompting them? Where do you think this goes in five years, like these just autonomous things that'll be just working for users directly? Oh man, Thomas, I think in like one month at a time, this stuff, this moves very quickly. Five years is difficult. I think it may be like a spectrum. So I think the lower bound is maybe a continuation of what we're seeing. And the thing that we're seeing right now is a giant speed up in engineers productivity. So within Anthropic, this is on the order of many hundreds of percentage points of productivity increase per engineer. And that productivity increase is actually accelerating. So every month it's bigger and bigger, despite the size of the engineering we're growing. So we're still constantly bottlenecked on engineers today, even though the productivity of every engineer is growing a ton. And this is what we're seeing with every customer that we have. So this is sort of the lower bound is that this trend continues. The upper bound is that the model is able to get so good at coding and so good at tool use and computer use. And this is exactly what we have been training it to do. The upper bound is that essentially software like disappears. And the same way that, you know, software replaced hardware, the model replaces software. And because anytime you need software, the model can just write it for you. I don't know where exactly it's going to be. And for Anthropic, the belief is it's somewhere in the middle or maybe one of, you know, maybe the extreme version. And our job is essentially to make it safe. So to make sure that the cyber risk is under control and, you know, like governments are bought in and we're advocating for the right policies and so on. But it seems inevitable that it's going to be somewhere between these. Yeah, I can tell you my personal view because I'm unbound from having to actually execute any of it. So I can just dream the dream and not have to worry about actually making it happen. But I think what you're going to see, like, for example, we know on this device, right, we get a software update and we have to download it and maybe fix some stuff that kind of worked before. I think that within, you know, five years, the whole concept of a software update will go away, right? Essentially, all the devices that we use will be self-healing. So same with my LG TV that's sitting on my screen here. Like it keeps prompting me, like every time I turn it on, I got to update the software and it's kind of this old, terrible software. Like take that for any device and it'll just go away. The software in real time will constantly be updating itself, constantly coming up with new features, constantly fixing itself to the point where it'll be completely transparent. You won't even think about a software update anymore. You'll just think about a service that you get. I turn on my TV and it does exactly what I want, right? I think we're going to see that replicated across almost every single device that we use. And it's the innovations that Boris and the team at Anthropic are doing, along with kind of the rest of the industry and the semis and all that stuff that are going to make that happen. So, well, look, Boris, we're very grateful. Thank you for taking a few minutes this morning to speak to us. I'll speak on behalf of Philippe and Kotu. We're incredibly proud that we're even associated in a small way with your career and obviously more largely with Anthropic as an investor. It's truly amazing what you've done. And I love that you think one month at a time. I can't wait to see what we'll be in a few months, but this is really exciting. So thank you. Thank you so much.