00:30:06 · Brian Armstrong, Molly O'Shea
[00:30:06] S0: Yeah, well, it's interesting. When I talk to our board members or even interns that have come into Coinbase, they all tell me this is one of the top few AI enabled companies in the world. Like, all companies right now are trying to figure out how to leverage AI, and and I think Coinbase is really on the frontier of this. So we've done that in a in a couple of different ways. One is we're using it to improve both the quantity and quality of code that we ship. The amount of code being shipped per developer is up about two x year over year. But believe it or not, you know, there's actually like, if the average developer is shipping maybe like eight pull requests a week, the top few engineers are doing like a 100 pull requests a week. So there's this actually like a true outlier effect of these like super builders that have become super AI enabled, and we're essentially taking those folks and using them to train the rest of the engineering team just to keep driving that number up. And then, yeah, I I I think, like, another thing we've been doing is we've actually been changing the traditional team size. Right? Traditionally, in tech companies, you'd have, like, a team of pod of 10. There'd be one product manager, one designer, eight engineers, something like that. And we're now seeing these teams that are two to four people. In some cases, we're seeing a one person team. [00:31:24] S1: Really?
00:31:24 · Molly O'Shea, Brian Armstrong
[00:31:24] S1: Really? [00:31:24] S0: Yeah. Because they're using 10 AI agents along with the one human. And so on many of these teams, we now have AI agents as collaborators. In the Slack channel, you can talk to them. They're creating pull requests. They're creating new designs. There's actually about 1,200 full time agents working at Coinbase now. Really? We try to basically estimate it because you can spin up an agent for five minutes and shut it down, but if you look at it as how much time AI agents are working compared to a typical forty to sixty hour work week, yeah, it's about 1,200 full time agents now work at Coinbase, and that's what's allowing us to get this productivity with higher quality, I should mention. We're actually seeing the rate of bugs and incidents go down per line of code shipped. And so it's driving up quality while driving up the quantity as well. [00:32:23] S1: I mean, there's mass hysteria for jobs right now. [00:32:26] S0: Yeah. [00:32:27] S1: How much of that do you think is rational versus, like, irrational? I mean, like, clearly speaking, organizations will be rebuilt around AI and technology and agents and token maxing for some, but what do you think the future organization would realistically take shape? Like if you were starting a company today?
00:32:54 · Brian Armstrong, Molly O'Shea
[00:32:54] S0: Yeah. So I think that the average team size will be smaller, but that just means companies can do more things. And so I don't actually, I'm in the camp of people who don't think AI is going to actually cause high unemployment. I I think that employment will be about the same. I think that work will be optional in some sense, in the longer term. I think most people will actually choose to work on something. It just might feel a little bit different. Right now, you're a podcaster, and I'm a CEO going and talking to people on stage. Like, a hundred years ago, this would not have been considered real work. Know? Like, real work is like digging a ditch in a field with like, most of the people in The US economy were working in agriculture with very manual labor, like a hundred years ago, and we don't have to do that anymore. So this is what we call work. I think that in the future, people are still gonna work on all kinds of things, but it may not be what we would consider work. AI is gonna eliminate tasks and toil. It's not gonna eliminate jobs, in my view. By the way, you reminded me of something else on your prior question about how we're shaping Coinbase. One of the next big things we're reaching for to get in place is is recursive self improvement. Right? [00:34:05] S1: So
00:34:05 · Molly O'Shea, Brian Armstrong
[00:34:05] S1: So [00:34:06] S0: we're actually using our apps to collect feedback from customers, whether that's bugs bugs they're encountering or even new feature requests. And we're using AI to aggregate all of this input from customers. And then the next set of agents actually take those priorities. They go plan it out. They draft the code and the pull requests. And then a human being can literally just sit there every day and say, okay, here's like the 100 things that we heard from customers. AI went and implemented them. You a can code review and say approve, approve, approve, approve. And like a 101 person is just reviewing a 100 changes that improvements that went out. And the next day the AI collects another set of 10,000 pieces of input from customers. So the closer we can get to this self recursive self improvement loop, I think we're gonna see really dramatic gains. [00:34:56] S1: Are you having fun with all of this?
00:34:56 · Molly O'Shea, Brian Armstrong
[00:34:56] S1: Are you having fun with all of this? [00:34:58] S0: Oh yeah. Mean, AI is making everything exciting right now. I think we're living in a golden age and, you know, it's part of the really exciting thing is we get to go build the financial infrastructure for this agentic economy. Most people, I think, today think of AI as like, oh, I just talk to one agent at a time, and maybe I ask it a question about my investment account or something. But in the future, you're increasingly gonna talk to one agent. That agent is gonna orchestrate hundreds of thousands of other agents. It's gonna pay for goods and services with companies that it needs to engage with to get work done on your behalf. And so, we need a new financial infrastructure for that. That can do very small amounts, very frequent amounts, where every AI agent can actually have its own financial account that doesn't have to go through a typical AI agents don't have a piece of government issued paper with their identity, and they can't fill out a CAPTCHA and all these things, so they're gonna need a new way to pay and have a financial account, and that's what we've built at Coinbase. [00:36:03] S1: How are you staying on top of everything? Max Levchin said that he's reading research papers. I interviewed Sebastian from Klarna. He's vibe coding and going on demos most of the time. How are you staying on track of everything on top of all the regulatory hot takes for crypto?
00:36:23 · Brian Armstrong
[00:36:23] S0: Yeah. Well, I mean, I have started writing some code again using AI agents. I test out the tools every few months, make sure I can ship a few pull requests and, like, still get in get in the weeds. But the other thing that I'm doing is I'm really just taking we have some of the best tech talent in the world, and some of these folks are I just go around and talk to all different teams. Some of them are creating a brain for their team where all the context of that team is getting updated continuously. So every time, like, a new lesson appears, whether from an AB test or customer feedback or an incident happens or anything like that, it's like the collective knowledge of that team just keeps getting updated in the brain of that team, right, in GitHub. Usually these are in markdown files. And we're gonna start to do that on more and more teams, and eventually you're gonna have be the intelligence of the whole company coming together, like the aggregate of all of these different team pieces of knowledge. Right? And that's what allows the AI agents to get smarter when they go build the next feature. And you have to it's like a shift in mindset where, you know, let's say that it creates a draft pull request for some new feature, and it's not quite perfect. Your instinct typically would be to go in there and actually edit the pull request to make it better. But what you wanna do is actually update the context, the brain that generated it, and say, okay, when you one of these, do it in this slightly different way. Don't forget about this. This is important, and have it rerun itself and generate a brand new one, and only when it one shots it perfectly, then do you ship it, because now you've encoded the thing that it missed in the brain for all future teammates and all future pull requests, and so it's a slight shift in how teams need to think about building the recursive self improvement of that brain over time, and not just editing the work of the AI itself.