Coinbase’s Everything Exchange: Agentic Finance, Stablecoins & Tokenization with CEO Brian Armstrong

The minimum fee typically on a debit or credit card transaction is about 30 cent fee flat plus the additional percentage. Arguably, it doesn't really work well for anything under a dollar. Certainly, it doesn't make sense to pay 30 cents to send a 1 cent payment. And I believe the stats we last saw on this are that about like 76% of the Agentic Commerce transactions we're seeing are under 30 cents. We don't want the AIs to be unbanked. You know, we want to bank the AIs. They deserve financial services as well. The AI agents themselves are going to have to have their own accounts. And so we've made some simple tools where with a single command, you just prompt, you paste it into your AI agent, it can now have its own financial account. So it'll be interesting to see how the agentic economy evolves. There'll be more agents than humans in the not too distant future. It kind of stands to reason that the agentic economy will be bigger than the human economy at some point in the not too distant future as well.

Yeah, good to see you, Elad.

Yeah.

Yeah, so we're at a really exciting time in Coinbase's history because on the one hand, every asset class is coming on-chain. It's now possible to trade everything in one place with good liquidity and good cross margin. And so for markets and trading and that, out there is stocks and commodities and crypto and derivatives, perpetual futures—like it's all coming to one place where traders can do everything in one spot, including prediction markets, too. And they can do that in more countries now given the regulatory clarity that's emerging, which we can talk about if you want. So the everything exchange is kind of our bread and butter, our core business. The next big area that's emerging is stablecoin payments, which have just been growing like crazy. Even when Bitcoin prices are down, stablecoin payments just keep growing. And so we're helping companies all over the world integrate stablecoin payments, which are fast and cheap and global. You can send money at the speed of information, right? It's kind of like under 1 second, under one cent US to move money anywhere in the world. And then the third big one is something I know you talk a lot about on this podcast, which is how is AI changing, because AI is changing everything and so it's obviously going to change financial services too. And so we call this Agentic Finance or AI-Fi. And we're building this into Coinbase in a number of different ways. The first way is that people who have their existing financial account, they're going to want to have an AI adviser. Kind of like how wealthy people are able to get access to these investment advisors. But everybody should have access to them now. And they can help teach you about building the right portfolio and dollar-cost averaging, tax-loss harvesting, and all these kind of things. And if the market dips 5 or 10 or 15%, leg into this kind of portfolio over time and try to save money on fees and earn yield while it's sitting in cash, and you can just talk to it like a human and it'll do all these things for you. At least that's the vision, and it's getting better and better. So we've launched an adviser in the Coinbase product. The next thing is that the AI agents themselves are going to have to have their own accounts, you know, and so we're on a mission to kind of go out there. We don't want the AIs to be unbanked, you know. We want to bank the AIs. They deserve financial services as well. And so we've made some simple tools where with a single command, you just prompt, you paste it into your AI agent, it can now have its own financial account and it's using a self-custodial wallet and crypto rails. So there's not like a KYC process. Our AI agents don't have a government ID. They can't walk into like a bank branch, at least not yet, without their humanoid robot companion. And so now any AI agent can have its own financial account to do payments, to pay for goods and—

Yeah, so we're doing it both ways actually, just because I think there's going to be demand for both use cases. So it's going to be possible to create a separate agentic account with segregated funds on Coinbase in the near future, and you can give your agent access to just that to do trading and payments on your behalf. That's kind of linked to your identity as a human. We're also going to provide self-custodial wallets to any AI agent out there that wants to hold a balance in stablecoin, or eventually it could actually raise money. It could issue a token. And so what's fascinating is that today we often talk to these agents and it's sort of like one person to one agent and they go get some work done, but they often get stuck. They get stuck. They're like, "Oh, here's a paywall. Can you put in your credit card? I need to spin up these AWS resources," or, "You wanted me to book this hotel for you, but here's the web page. Can you go put in your credit card?" That kind of thing. And in the future, I think you're going to want to have these agents have their own spend accounts, kind of like having a personal assistant or an employee who has their own spend account. It's going to need probably like a stablecoin-backed credit card for backward compatibility, but we're seeing a lot of agentic payment now flow over these new protocols like X42, which is one that we incubated at Coinbase that we actually put into the Linux Foundation. So, Google and Cloudflare and AWS, a bunch of people are working with us on that now. And so, that's been really cool. And we put together like this little fun website called agentic.market where people can go see the stats of how many payments are flowing in this new economy and see all the different services. So we're helping a lot of businesses get enabled into the agentic economy, like, "Is your store ready to accept agents as customers?" We have a product called Coinbase Business that helps companies accept stablecoin payments, and that's good for humans who want to pay with stablecoin payments, but it's also good for agents, and they kind of get both out of the box.

Yeah. So what's interesting is a lot of the agentic commerce that's happening, these are pretty small transaction sizes. Sometimes just a few cents, and as you probably know, the minimum fee typically on a debit or credit card transaction is like a 30-cent fee flat plus the additional percentage. So arguably it doesn't really work well for anything under a dollar. Certainly anything under 30 cents. Like it doesn't make sense to pay 30 cents to send a 1-cent payment. And I believe the stats we last saw on this are that about like 76% of the Agentic Commerce transactions we're seeing are under 30 cents.

So a lot of it is gathering information. And so you could think of it like it's venture firms doing research compiling from paywalled data, or it's recruiters who need to scrape data from LinkedIn. It's oftentimes actually agents talking to other agents that are specialized in some way and getting source data from them, almost like a tool call that's happening. And on that agentic.market website, we have a list of the top service providers that are doing like search and financial data. And I think what'll be interesting is in the future, you'll see actually other agents in the market that become specialized. The frontier models are trained to be good generalists, right? And they're kind of good at a lot of things. And we're seeing instances where smaller models that are open-weight models can actually be fine-tuned, or through reinforcement learning they can outperform the frontier models on a specialized task within the company, right? Like I'm sure you've talked to lots of startups about this kind of thing, but it's like a compliance review queue, where if you train it on 100,000 compliance cases within Coinbase proprietary data, a small open-weight model can actually outperform a frontier model. And so I think you're going to see in this marketplace a bunch of specialist agents. It might be like, "Hey, this agent is just really good at ordering you pizza if that's what you want." Or this agent is like a really incredible designer and that's all it's been trained on is like tons of proprietary Figma data or something like that. So it'll be interesting to see how the agentic economy evolves. But you know, it's probably common wisdom at this point amongst your listeners, but there'll be more agents than humans in the not too distant future. It kind of stands to reason that the agentic economy will be bigger than the human economy at some point in the not too distant future as well.

Yeah. Well, just actually on that, it's funny. I've heard people like Elon say that they think there won't be money in the future, right? And I actually agree with Elon on most things, but that one I was like, okay, I think I know what he means. He's basically saying at some point you start to measure things in energy and mass, and like those become the scarce resource. And I think in the end, like in the limit case, like he has a point. But the thing is there's going to be scarcity for a long time. Even if the cost of goods and services really falls, like we have price deflation because robots and AI start to just make commodities and food and housing very cheap—that could happen. The government printing might also create a bunch of inflationary pressure. But there's always still going to be scarcity of certain things, like land, especially usable land in various ways, energy, like chips. Dyson spheres are going to be expensive, I think, pretty far into the future. So I think we're still going to need a medium of exchange. And it's really cool to see that as AI emerges, crypto rails are just kind of a nice complement for all that stuff to finally work. Like crypto was, I think, really, really good for humans and it's going to be essential for AI.

Yeah. Well, I'm reminded of that meme. Have you seen that meme that says, "¿Por qué no los dos?"

Yeah, it seems like they just might.

Yeah. So like a lot of companies, we started off with the basics. You know, we got some coding tools, we got it hooked up into our various data repositories. We've been using it for a long time on machine learning for risk prevention, fraud prevention. Customer support inquiries started going through AI. So that was all like good table stakes stuff. And then the things that we've started to do recently, where we're sort of pushing the boundaries here, I think one of them is that we basically want to get to as close as we can and actually achieve recursive self-improvement, right? And what does that mean in various ways? It means that you have a brain for the company. That can be a whole bunch of subsets, by the way. You can have a brain for each team, you can have a brain for each service repository within the company, you can have a brain for each individual, which, by the way, I believe they should be able to take with them when they leave Coinbase. So we're creating some good policies on that. Let's say you have a service like a GitHub repository or some backend service that's running, and today we have engineers spinning up agents to make changes to that. What we're just now launching and seeing recently is every time the agent comes in to make a change to that service, it's actually ingesting the brain for the team working on it and for that repository. And so what is that brain? It's basically a history of all the incidents that have ever happened with that service, all of the financial controls that have to be enforced with that, every A/B test that we've ever run on that, of course like the full GitHub history of pull requests that were accepted or rejected. And so the people running that team at any time can kind of go in there and look at it. Today it's just markdown files, right? It sounds kind of simple at the end of the day. But getting this all hooked up—so let's say you ask the agent, "Hey, I want to make a change to this service." It ingests the brain, it makes that change, and then as a human you go to review it and you're like, "Ah, it missed something." Now the simple thing someone might do in that moment is go make a manual edit to fix it and then ship it. But the key step that we're trying to really enforce now in our software factory or software development lifecycle is when you go to make that change and edit the agent's thinking, that context has to go back into the brain so that you not only fix it in this case, but in all future cases going forward. And then basically what you start to see is the accept rate for one-shotted PRs or that kind of thing just starts to tick up over time, and you've now got like a recursive self-improvement system. So that's one of the big ideas we've been working on recently, and just getting our own internal agent harness, which we call Toshi, just integrated with all these tools. Like it can now make calls out to external vendors and pay using the agentic finance economy over these X42 protocols. If you want to edit the brain as you're making a change to that PR and it goes back into the brain, putting that into the harness. Actually just earlier today I was making a plan for this relatively complex feature. It was kind of amazing. It's now even allowing people like me, the CEO, to get back in there and ship features. But I asked one of the high-end, expensive models, "Okay, go make a plan for how this feature might be implemented. Break it down into three phases. Each phase has 10 different pieces that need to get built." And I was like, "Okay, go execute phase one. Go spin up 10 separate agents to go execute each of those pieces in parallel. And by the way, recommend the right agent to use because they might be cheaper models." And so it spun up a bunch of open source, it spun up a bunch of Grok. And just two minutes before I hopped on this podcast, it just pinged me and was like, "By the way, all 10 of those things from phase one are now done and ready for you to review." So, anyways, I'm starting to get to this place where it's kind of addictive, actually, because in the past every time I'd have these ideas or bugs I find, I'm just pinging the team all day long like, "Hey, can we do this? Can we do this? Can we do this?" And now instead of pinging the team in Slack, I can basically tag an agent in Slack with that comment or drop it into our internal harness, and instead of bugging the team and asking to do something, I just send them a PR for review and it's already done. And that's a magical moment that is very addictive.

Well, productivity is definitely increasing, right? So, that's slightly different though than saying the average company size will be smaller. It's kind of like just saying tasks are being automated or eliminated. People are not being eliminated; I think tasks are being eliminated. And so if you take a company like any sort of company today, Coinbase or others that exist today, it's not that I think we're going to shrink down the headcount to 10 people or something like that. I think that the existing people we have are going to be able to just get much more done and the pace of everything will accelerate. Now, it's also true you're going to see some companies that are two people or five people or whatever and they're able to do a lot more than what we would have historically thought was possible. And that's also true.