Should we give AI a bank account?

But you said earlier that most economic transactions in the future will be agent to agent. And you said this is something we'll want. Why do we really want that for all the people who are listening being like, that sounds insane. I think it's not even the most. I think it's all, I think it's every single transaction will be executed by, by AI in the future. You know, it sounds a little science fiction, but that's the world. I think we will ultimately want to get to, if we have the proper guardrails and it's sort of aligned with human ethics and so on. I hope it's not just a chat interface on top of a shopping cart. I hope it's not just a new sort of search that still has the same underlying mechanics. You know, agents don't need shopping carts. Will we see a single person company that can generate a billion dollars or a hundred billion dollars value? And my response to that always, why do we need that one person? Most people think about AI as something that helps us work faster. Sean Neville is building for a different future. One where AI doesn't just assist with finance, it is the customer. Sean co-founded Circle, an architect in US DC, the 76 billion dollar stable coin that turned crypto from speculation into actual infrastructure, but even inside Circle. He says not many people believed in it at first. Now he's doing it again. Catana Labs is building the first AI native bank, not a bank that uses AI tools, but a bank designed from the ground up, assuming your customer might be an autonomous agent that exists for minutes, executes a transaction and disappears. This raises questions we've never had to answer before. How do you do KYC on something that doesn't have a Social Security number? Who's liable when an agent misbehaves? And why do credit cards fail where stable coins succeed when AI is the one spending the money? Sean has a hundred AI agents doing compliance monitoring right now, overseen by just two humans. He's navigating regulators. We either want to build frameworks for AI finance or believe nothing good can possibly come from letting machines touch money. This conversation is about building at the collision point of two platform shifts and why some infrastructure has to be rebuilt from scratch. Welcome to possible, Sean Neville. Great to have you here. Welcome to possible. Thanks. So you helped start a company called Circle, which helps process, I think now 10 trillion in volume using stable coins and valued at close to 13 billion. Let's rewind to 2013. You and Jeremy co-founded Circle with this bold vision of making money move like the internet. What was the moment or idea that convinced you this needed to exist? And what made you guys go, we're the right people to build it? Well, so this is actually the third company that Jeremy and I worked on together. And so our previous companies were at the advent of the web, where we saw open standards on the internet and able people to publish their opinions, express themselves without gatekeepers. So fundamentally, when we look at Circle, that's really the seminal idea is it should be possible for everyone to plug into the global economy without gatekeepers sort of chopping off your money along the way for the privilege of access. And so the vision, just generally speaking, was how do we make money work on the internet the way content and data work on the internet, where the internet has new business models, but it wants data transmission to be free. And we view money as just data. So transmission of money should be free or very, very close to free so that we can unlock new kinds of businesses for everybody. And so that was generally the idea. Why us? You know, we had a lot of experience in building web companies at that point and working on sort of open standard approaches to it. But we were also hopefully naive about doing that in a regulated industry. This is a regulatory pattern for the internet. It's quite different than obviously financial services, but also not only did we not have a background of, say, payments, traditional payments, but really any, we were builders in any regulated like healthcare or aviation or you know, pick your regulated industry. So I think it was actually helpful that we were so naive that we thought this is the thing that we could, that we could do on the policy side in addition to the, in addition to the tech side. What was the regulatory thing that was kind of the most interesting, like a surprise or a learning or something you'd say to other entrepreneurs? Cause I, by the way, I've also had similar kind of things, both in financial systems and in biology and pharma, where I go, ooh, I'm like, I go from being a software guy into it. I'm curious which one is the most the, the aha or the moment in discovery and regulatory. Yeah, it was both, uh, and aha, at the same time. And I think, you know, we embraced the opportunity to educate policymakers and we felt like even if it took a decade or more to encode some of these ideas and public policy, we were up for that. And that's what we felt needed to happen. I mean, a lot of the early economic philosophy, particularly from Jeremy, you know, my co-founder, and then I embraced as well was just this notion of what's been referred to as a narrow bank, which is this idea that's been around since the twenties or thirties, um, which is a really simple, almost boringly simple idea, which is just, you know, if you're a bank and I give you my money and I come back 90 days later, my money is still there and you can give it back to me because the sort of like payment facility, custody facility is separate from the credit facility of a bank and you're not allowed to, you know, take my $10 and lend it out a hundred times. And that's the thing that hasn't actually been allowed, um, in the government and stablecoins are effectively that they're narrow banking. Yes, exactly. Now, and one of the cleverest things I think we did in the regulatory side and PayPal, which was, Hey, if we could get the FDIC to say that our deposits are FDIC insured, you basically get all the regulators go, okay, fine. Like we're good. And so what we did is we essentially went to them and said, well, you have this pass through broker insurance, like where it allows it. So we're not, we're putting it in banks on behalf of the people. So we're put, but, but it's FDIC insured still and the FDIC agreed with us. So then our regulatory conversation got a lot simpler because anytime the regular asked us, well, but you realized you're protected by the FDIC and they're like, Oh, nevermind. So that was part of it. So Sean, for our listeners who might not be as familiar with the stablecoin, could you explain to us what it is? Yeah, I don't love the word stablecoin. I think of it as dollars and money on the internet. And, and, you know, I don't know if the word stablecoins survives, but the origin of it is that we have other kinds of money that can move on these open rails, these open blockchain rails. Bitcoin, for instance, is one and there are many others. But the problem with those coins is that they change in value so that if I'm, if I'm sending you money, how, how much money you receive might be quite different than the money that I sent. And neither one of us can know ahead of time. And so the idea of a stablecoin is that I send you $10 and you get $10. Um, and it's the stable in terms of its value. So, um, before you get to the really exciting stuff you're doing now, open claw went from zero to I think over 145,000 getup stars in two weeks, 1.5 million agents, transacting on mold book. Some people are calling it the early singularity. I'm not one of those. Others are calling it a dumpster fire. I'm closer to that. Where do you land? Maybe both in a way. Um, so I think the positives I would say is that very quickly, many more people have suddenly experienced this idea of a layer of abstraction above, say, Claude code or chat or co-work, even, you know, these new, uh, sort of what they had thought of is how you interact with the, uh, sort of a layer above that they can use all of those things under the covers. And so the appetite is there. Um, and you know, the interest and enthusiasm for those kinds of actors is there. More on the dumpster fire side is that, um, you know, they're, they're just tremendous security implications. Uh, I think that, you know, the top open claw skills are, I have malware problems. Um, people are using them without being able to sort of clear the trust chasm. And when it comes to those agents being financial actors, that's a real problem because people can get hurt. Uh, and it could go sideways, uh, in an ugly way, which is many of the things we want to address and we want to make it safe for agents to be economic actors. I mean, ultimately, we think the only economic actors anyone will trust will be agents, but it's not. Maltz or lobsters or whatever you sort of call them today. So I think the experiment is great, but there's, there's also a lot of reasons to be worried about the form, a particular form that it's taken right now. So actually that leads perfectly into my next question. Sort of talking about how do you make economic agents trust worthy on the internet? You stepped back from circle, the company you co-founded and started a new company. Can you tell us about it? Yeah. So in many ways it's a, it's a nice trajectory from circle and, and I have to, you know, I am still on the boards. I'll bite my tongue on some things on the, the circle side, but you know, we got excited about creating dollars on the internet to unlock new business opportunities. But one of the other problems with the economy is that people don't have equal access to it. And one of the things that's exciting about AI is that it's, if we get it right, then the optimistic view is it will be able to provide individuals and businesses everywhere around the world with access they've never seen before. And the economics will work and the intelligence will work. And, you know, it will leverage things like digital dollars and other currencies on the internet, but it'll do more than that. It'll, it'll allow people to take the next step into, you know, stepping into the, the global economy. They just haven't been able to before. And so it's a nice continuation, but it's also the thing I've been, you know, most excited about in my life to work on. It's, this is just a, this is an incredible area to be, to be building. So walk us through a concrete example. Uh, an AI agent tries to pay for something today using existing rails. What breaks? What does, where does a credit card fail? Uh, where a stable coin or, um, other, um, rails succeed? Yeah. So I, I'll give a, uh, it's kind of a retail commerce example, even though most, the flows that we're focused on now are actually more, uh, in the enterprise and business oriented, but the commerce one is almost easier to, I think, um, illustrate. Um, so it starts with, um, maybe an analogy to how you might buy something today. So if you were purchasing something on Amazon.com today, we don't have to really worry that we're on, that we're not on Amazon because SSL and HTTPS and all that. So we don't, we used to have locks in the browser that would at least give, now we don't even need that. We just trusted the technology handles that we're talking to the legit Amazon. Um, and that kind of layer doesn't exist with agents today. If they're, if they're chatbots on a website, it exists. But with these new surfaces, how do I know that it's Amazon that my agent is talking to? So this is the fundamental identity issue, um, mapped back to an institution that I want to have some trust in. Um, and then once I clear that hurdle and I can be sure, uh, that my agent is talking to Reed's agent, how do I put some controls around what those agents are allowed to do? Um, and a simple example would be spending rules. If just set up something like allow my agent to buy, um, you know, my goods on Amazon up to $50, but notify me for anything above that and don't let it spend on Etsy without notify me. It was simple like spend rules and they can get much more complicated, but just versus identification. And then there's rules and then there's the auditability of looking back and seeing what actually happened. And that's particularly important in cases where something went wrong, uh, and understanding where the liability lies. And so these are all these sort of, you know, who are you? What are you allowed to do? And can you look back and understand what was done? Um, those are sort of the initial fundamental problems that need to be solved if agents are the ones that are, whether they're paying other agents or paying humans or just, uh, participating in automation with finance. It's, it's really all of these things. So loosely we've referred to these things as KYA, which is know your agent. Um, and you know, being aware of who you're interacting with. And if it's my agent, how do I make sure it's trustworthy to your agent? You know, these sorts of things. And so this is kind of a foundational element before we even get the payments. That I think, you know, we need to address that we've been focused on. Hopefully the solution is not a proprietary vendor solution, but something that is more of a de facto or, or, or formal standard that we can all build on. Similar to what we saw in Lockheed commerce. So what are some of the things that go into this new architecture? Obviously agent identity and certification as one part of it. I'm curious to get a little bit more depth. What are the different kinds of ways that you conceptualize the way the network works when the customers or the engagement services are AI, not humans. So, you know, there needs to be a way, uh, and this, this is like nascent. It doesn't really exist today. There isn't sort of DNS of agents to just resolve to like discover agents out there. Although there are things that are emerging related to MCP server marketplaces, for instance, or now we have, you know, open claw skills marketplaces. Um, but, um, there isn't that mechanism yet. So, um, one of the things that needs to happen is, uh, when agents begin to talk to each other, whether it's through API, is MCP servers or through these new discovery mechanisms, the first, some verification can happen that maps back to who is the operator, uh, of that agent. It could, if it's a one to one scenario, you have your agent and I just need to be sure it's actually your agent, but it could be much more complex and that we're all using someone else's agent to perform some function. And then we need to be sure they operate or that agent is who we think in the regulated financial situation. Um, it often needs to map back to a licensed entity that's allowed to execute money transmission or lending or, or whatever it may be. And so, um, we think of just the foundational primitives is a lot of them are related to identity and what are the protocols meaning? Just how do you exchange these credentials back and forth, uh, in order for agents to do that in a trustworthy way. So Sean, at Circle, you were designing APIs for human customers at Katana. As you just said, you were designing for AI customers. And so can you give us another concrete example of like, what's different? What are API design choices that you're making now because you're specifically building for these AI agents? Yeah. I mean, I think, well, first of all, I think API design is just totally different now. Um, you know, as a, as a human developer, do you ever really need to read docs anymore to integrate in API, whether you're building something in finance or not. So a lot of, you know, even the API docs were generated for agentic workflows or even just directly all alims to be able to parse effectively. Um, and you know, many times this is sort of even past the era of vibe coders, but, but, um, you know, principal architects leveraging large language models to build software don't even look at the underlying APIs anymore because it's more productive. So long as you can trust the AI, it's more productive not to do so. So then you end up designing APIs in different, in a different fashion. Um, there's a really simple version of this that used to happen prior to some of the latest advancements, which was just how do you get in a large language model to use a tool, which is really like, if you put in these words, how do you get that large language model to recognize, okay, now I need to make an API call. And so there are these function calling sort of tools, mechanisms. And even in that mode, you could see in order to tweak the large language model to understand it should use an API. You kind of had to do some prompt prompt engineering and coax it a certain way. But the real way to make it more reliable is actually change your API. So that the, particularly not that so much the signatures, they get a little technical, but that's so much the function signatures, but the input and output parameters. So that they could understand the schemas and what it was meant to deliver. You would end up changing in a way that was tuned to large language models, not so much human developers, it would be looking at like a JSON schema to read. And that is only becoming more exacerbated. So I think that's generally speaking, the case to the circle question. I think, you know, when we build APIs and sort of help businesses leverage stable coins with APIs, we're optimizing to make them productive and make them happy, make developers happy with AI. There's a little more of a governance criteria in place, which is how do we make this AI trustworthy? And how do we, how do we help this API be reliably and safely executed by this machine? And it's a slight difference because trust is still the biggest hurdle to clear. But with all of, how can we trust these things with anything, let alone money? So looking forward a little bit, where do you think the first large scale adoption of AI transacting on behalf of consumers, enterprises, are there any obvious industries that might be disrupted or transformed? So I think ultimately it will hit everything, but obviously not all at once. I do have very strong conviction that, you know, as I said, I think the only actors that we'll trust with our money will be agents and agentic actors. And it'll be the only competitive way to earn a return and so on. But we're certainly not there yet. And so the harder question is, well, what, what does it hit first? And many of the scenarios that have been outlined, I would say are more consumer oriented, but where we see flows happening today are much more in the enterprise. And I wouldn't necessarily call them agents even. They're sort of automated workflows that increasingly have large language models in the loop. And in the world of finance, they're not, they're not sort of like the sexy use cases of get me some edge and, you know, predict the stock market. And it's more like the disappearing back office. It's cash flow, you know, basic liquidity management, compliance, reconciliation, you know, these sorts of things. And so, so I think the, what happens sooner rather than later, doesn't really look like a trading floor full of agents. It looks like a back office that's gone in the enterprise. So I was trying to explain to my 10 year old the other day how people used to buy airline tickets and I couldn't remember. I was like, God, what did we do? Did we call someone like before there's the internet? When you think of the consumer side, how do you think consumers will sort of act differently? Like, obviously we went from like, we can't trust the internet for anything. Oh, we can buy books. I guess we can buy clothes. Now we can, you know, buy cars, buy whatever. Is there going to be like a similar trajectory? Yeah, buy stablecoins. Like just for the sort of normie person who does online banking or puts their credit card into Amazon, like, what are they going to expect from this, you know, sort of AI agent future? Well, I hope it's not just a chat interface on top of a shopping cart. Yeah. You know, I hope it's not just a new sort of search that still has the same underlying mechanics. You know, agents don't need shopping carts and we don't need, you know, user experience sort of, here's how your tax is and like they can understand a lot of that data effectively. So I think the exciting element for consumers will really be expecting a new kind of interface that does not look like e-commerce that optimizes you to not drop off in a shopping cart experience across a number of pages. And what does that experience look like? There's a lot of different experiments there. And so I'm not going to predict exactly what it looks like. I think it will look very, if we get it right again, I think we'll look very, very different than today, but it won't, it also won't be the case that an agent just magically handles all of your payments because there is some kind of shopping that is just pleasurable and requires sort of an inner sort of introduction of taste. And then there are other sort of transactions for commerce transactions that actually I never want to have to buy a train ticket again. I just wanted to handle that for me. But I think, you know, there are all of these little pieces to resolve. I think the user experience and the consumer experience is the hardest not to cry. Gotcha. So getting back to the trust piece, obviously we have agent native banking needs like a new compliance strategy. Like that's going to be hugely important. And so can you tell us a bit how Cotena handles compliance monitoring and like what do humans do versus what do AI agents handle? Yeah. So I'd say first of all, being a regulated financial institution, we're subject to the regulations that exist. So we can hope to encode into policy additional things related to AI safety, particularly AI safety related to financial flows, which I think ultimately will become very important. But in the meantime, there are regulations that exist. So we're not starting from a blank slate. And so, you know, we police for money laundering, for instance, maybe put it in two buckets, so it's sort of real time risk monitoring for things. And we built out these systems, obviously at Circle. When you take a fiat money, you take a credit card and you want to turn that into stable coins, you need to make sure that the risk is managed effectively. And that's often handled in real time. And some compliance pieces are also handled in real time, like sort of screening, hey, this is a counterparty that we're just not allowed as a U.S. company to send money to. And it's just like a hard line. You can't do that. O-FAC screening. Exactly. And so, you know, we can't facilitate money transmission to North Korea, for instance. And that's just a hard line as a U.S. company. There are other things that are more complicated on the compliance side that usually not handled in real time. And a lot of that is relates to looking at transaction activity to understand, you know, money laundering schemes. And the traditional way to handle that is to have a bunch of humans who use tools, they sort of sign into like a SaaS like dashboard. And they do transaction monitoring and they open cases and gather evidence and file suspicious activity reports, which is like a very sort of secretive process. And a lot of that can be automated very, very effectively. And it doesn't eliminate the need for a human, you know, compliance and risk expert, but the tasks that that human does are very different when they're not signing into all these dashboards to do the work. They're more orchestrating agents that can very effectively do that kind of work. And how close are you to that team of agents doing this work today? Well, we do that today. And it'll only get better is what I say, you know, and we'll quantify along the way, is it getting better? Because these things are very measurable in terms of false positives. So we do that today. I think there's still an open question though about how you structure the team to orchestrate them effectively and the mechanisms to orchestrate them are changing so quickly that the answer today probably isn't the answer tomorrow. And I, but I think that's true across many disciplines. We're talking about compliance and risk, but it's also true in every other discipline when we're building this kind of, we're building an AI native bank from the ground up. And so it touches every operational aspect of the institution. Makes sense. So speaking about identity, obviously their agents don't have social security numbers. They don't have credit histories. Like when you think about traditional banking, it's KYC, know your customer, AML, anti-money laundering. Like how do you know your customer when your customer is an agent? So in the near term, our version of KYA is mapping back to one of those entities that has been KYC and KYB because an agent can't, it can't apply for a bank account. It can't, it can't get a credit card directly. And so in the future, I think that that will change. I think we will want it to change. And like we collectively, I think we'll want it to change because the value that we stand to gain will be great enough to force those sort of policy changes where there will be unique identifiers that are recognized by governments for agents. And even if the agent isn't identified with a specific human person. Yes. Okay. I think we will want it to get there because what we'll want to see ultimately is, you know, sort of really autonomous operations that many of us can kind of own and participate in and generate value from. And so at that point, the agent's identity really is separate from any one of us. But in the near term, it's more like, if you're interacting with our AI bank, then it needs to map back to us as the licensed financial, and I'm using Bank Loosely here, licensed financial institution, because an agent simply can't be a licensed financial, it can't go out and get money transmission licenses that are apply for a charter, you know, and so on. So there's a progression to get there. We have the technology to very reliably map back to the licensed entity. So the KYB business that's on board or the KYC individual, we can use cryptography to do that in a way that means you don't really have to trust any business or company to see the tie. And so, you know, we can see back to, like, say this merchant has these credentials that were issued by Visa to this business, even though we're only interacting with that merchant's agent as an example. You mentioned in the future that we might not want one agent to track directly to one human. It might be an agent that's acting on behalf of many people and they sort of have some ownership. Like, is there a concrete example that you can give where we might want that sort of shared ownership of an agent? Yeah. So I'll answer a different question first and then maybe that'll help answer that question. So there's, you know, there's been a lot of talk about will we see a single person company that can generate a billion dollars or a hundred billion dollars value? And my response to that always, why do we need that one person? You're like, forget the one. Let's do zero. I mean, really. So I think what we want is a different kind of relationship between the humans and the companies, the entities. And I think it looks a little different than classic equity, but there is a role for humans that are collectively orchestrating data and behaviors in an autonomous workflow or agent for them to generate value for themselves, mapped to the value that they're adding to the agent. And in that situation, you want multiple people contributing, not just one sort of orchestrating mastermind running a company that has a bunch of like co-pilot workers. So, you know, it sounds a little science fiction, but that's, you know, that's the world I think we will ultimately want to get to if we have the proper guardrails and it's sort of aligned with human ethics and so on, all these important issues. And then the identity piece becomes interesting because it's bi-directional between humans and agents. It's not only can my agent trust Reed's agent, it's more can Reed's agent trust me as a human being when it begins to interact with these workflows. And so then we get to this whole other level of, you know, what this could look like. This is more 10, 15 years, although now 10, 15 years, maybe that's three years. But even 10 isn't like that far in the future. And I think also to your point, like everyone's always asking, can we trust agents? Like, there's a lot of non-trustworthy humans around. And so if we could have a situation where we trust agents, that might be a better future. Yes. And, you know, the time is always the hardest thing for me to predict on this. And we're in this world and we see this constantly. But then, you know, if I interact with, you know, others who were not in this world, you could see this might take a little bit longer than we think to fully, this is the same thing that we went through with the internet and certainly with stablecoins. And now we're 13 years into circles tenure. And we're only just now realizing some of the use cases we felt surely will be satisfied by, you know, 2014. And they still haven't been. Right. And so it may take longer than we expect. Although right now this world surely seems, you know, to me seems very different than anything else I've done. In May 2025, you released the agent commerce kit. I do think the acronym is a little odd hack. But the month before, Google's AP2 or OpenAI is ACP. And then OpenCALL happened. All of a sudden, the future that you're working towards with agent identity, payment rolls, et cetera, suddenly is like, you know, kind of an interesting shape, suddenly much closer. What do you think this, with this kick off of this open source hobby project, accelerating everything and what are the ways that we need to navigate? What are the things we need to be doing in society given this acceleration of this? I do think that the solution for agentic identity and verifications and all the things we've been talking about should be something that is an open standard that everyone can build on. I'm a very big believer in open standards and the power of open networks, as opposed to say, you know, proprietary solutions offered by one vendor. And I think that there has been fragmentation on the AI side when it comes to these kinds of topics today. For a lot of reasons, I think one of the reasons is that even large incumbents aren't sure where the value is going to accrue, especially now that TAC is not a moat for anybody, one person or, you know, Alibaba. It's sort of no longer where I mean, maybe past the castles of moats metaphor and we're in the era of like gunpowder and ballista and, you know. So I think it's, for that reason and many others, the approaches to things like agentic identity are quite fragmented. They're even more fragmented now than they were a year ago. And so it feels like we haven't made the progress toward a standard that I had hoped for. But there are many ways to, for standards to emerge. You know, MCP has kind of become a de facto standard. It didn't emerge through the W3C as a working group, but competitors to Anthropoc have embraced it. You know, these projects that achieve a level of virality show real appetite and interest that can shine a spotlight on a use case that causes others to act. Agents suddenly showed up as customers a lot sooner than we expected them to. So that's a positive way to look at it. For ACK and for some of these other protocols, I fundamentally look at it as, you know, ideally a set of standards that we can all rely on so that as you're building agents and others are building agents, they're all leveraging the same open standard that no one can sort of unilaterally change. And also so that our agents don't have to try to integrate 20 different approaches to the same thing. And we can build just much more valuable competitive offerings on top of that base. So when people are worried about AI, they talk about runaway agents and sometimes that's, you know, sort of more bio terrorism and geopolitics. But often it's, you know, an agent's buying 10,000 burritos or, you know, making a million dollar poly market bet when you wanted them to put, you know, $10 on red. And so you talked about sort of guardrails that's like, you can tell them, you know, don't spend more than $50 or these are the rules. Talk more about your guardrails and specifically whether they're sort of rules or judgment. Like, are we giving these agents values about what to do? Or are we giving them hard and fast rules to make sure that they still have their autonomy but they're aligned with humans? Yeah. So it's a really good question. I think of it in two layers and this is how we've approached what we build out today. The first layer is very much, you could just call it the AI layer. And a simple example of a guardrail there, it would be easy to circumvent, but it's just kind of a prompt suggestion. But the problem with guardrails without policy enforcement is it's just that it just becomes suggestions. So you can flesh out many things, many layers of guardrails. You can have, we're talking about an agent, but usually there are many sort of LLMs involved in workflows and you can have LLMs check the output of other LLMs and did they have just once dedicated to guardrails before any activity happens? And so there is a sophisticated tuning of that, you know, sort of architecture and orchestration. But when it comes to things like moving money, those are deterministic outcomes and not, they should not be completely non-deterministic. If I'm moving a million dollars, it either needs to move or it needs to not move. And it can't be sort of, you know, worse than, then sort of failed. It can't say, oh, yes, it did move and it didn't. It moved to a different place, you know, hallucinations or whatever. And so the second layer is really the underlying deterministic layer, which, you know, is a place where stablecoins actually excel because we can programmatically enforce rules, both at the wallet sort of cryptographic key level and on chain and smart contracts. We can write rules that AI can't possibly circumvent. And so you need both operating and concert. And so that's sort of how we approach, you know, enforcement of these guardrails at a couple of different levels. And then I think there'll be iterations on both of these things, but it's one of those places where, you know, blockchains are useful for moving value, but the thing that's particularly exciting is you can program money on them. And you can write these so-called smart contracts that have rules in them that then are not controlled inside sort of a bank's ledger, but publicly on these rails using cryptography. And so that turns out to be really helpful for AI flows and AI actors who are executing these money flows, as opposed to just having an API and then trying to deal with the inputs and outputs. So another thing people are really worried about, of course, is external threats. And like, is this, are the AI agents going to become so good at whatever it is? Scams, fraud, etc. Like, are there ways that you think you are better positioned to ward off external threats than traditional banks? Or how do you guys look at that? Yeah, it's a problem now. I think a lot of people are approaching through this sort of proof of personhood, which is very useful. And it will become incredibly more useful. But we really think more of the other side, which is sort of proof of agency, which, you know, is more along the lines of proof that this agent belongs to the operator that it says it does. And that has these policies that we can be very sure cannot be bypassed. And how do we enforce these restrictions? And I think that, you know, we have a time window here to get that right, but the window won't be open forever before people are burned. And so some of the, the sort of not doomsday, but dire scenarios you mentioned, I think, are likely to happen without these guardrails in place, frankly. Absolutely. And some of earlier comments, you've mentioned that agents are surprisingly bad at marketing content generation, but great at compliance. What other counterintuitive things have you learned about what AI is, is actually good at and actually still TBD in financial services? You know, some of the things that it used to not be so great at, it's actually, it's not bad at now. And there's always the thing of, you know, I hear people say, well, I put this in Claude and it told me this or Chat, you can tear or whatever. And it's like, well, that's what it told you based on what you put in and what you have, you know, it's like sort of saying, well, I put this in Photoshop and this is the image that came out. I was like, well, if I use Photoshop, it's quite different than if a real photographer. Maybe you're the problem here. Yeah. It's like, you know, it's some of its garbage in garbage out and, but it's also, we have very sophisticated workflows now in terms of managing memory and context and so on, when it comes to making these things good at software engineering that have made them so good at developing software that the profession is completely different than it was a year ago. And I think that'll happen to maybe not every knowledge worker domain all at once, but it'll happen to many of them and it'll be a sort of, you know, similar progression. So a lot of the things that marketing content, in just in terms of generating copy, there's a certain tone and it's not just the m dashes or whatever the cliches are. It's the phrasing of the, you know, short, short, long, short, short. So that kind of like, you know, freshman in high school sort of like, this is a good essay kind of a thing, but that's going away with the right context and the right sort of, you know, data provided and the examples provided and models that get sort of teach each other to be better at those things. So, you know, I think the other answer to the question is a lot of the stuff that it's good at, it's the non-sexy stuff. It's just really good at parsing a tremendous amount of information, synthesizing it and turning it around to make you more productive. Well, it's so clear that you are so excited about this sort of next world and you're, you know, building this whole fleet of financial tools for AI agents. What made you decide that like this was the thing to do? This was the first thing you needed to get out into the world. Yeah, I mean, I think if we don't get this piece out, it's like an unlock. If we don't get this piece out, then we can unlock these new opportunities. I mean, it's sort of patterned through, you know, crypto and stablecoins and circle. I think it's relevant to AI is there's kind of a toy phase, you know, where people are speculating with some random stuff and it seems easy to dismiss, but gets some attention. But then there's this like really hard middle ground of infrastructure development and regulators start plugging in. And this middle ground can last a long time. I was doing really hard work to unlock new kinds of opportunities to the later phase, which is these like world changing applications, the likes of which we can't possibly imagine when we're in these early stages. And it was similar to, you know, when I was actually young, you know, in the early days of the web, could have possibly predicted all of the things that the internet and the web would be used for, but we just had a sense this is world changing. You know, that's the excitement is to build the unlock. We're in the infrastructure phase and it's often accompanied, if it's truly valuable, it is often accompanied by a speculative element as well. So there's hype and there's speculative cycles, but that doesn't mean it's not real and important. And so, and the last thing I would say is I think that there's a time window here to get this right and the window is closing. And so there's urgency here to execute. Why do you think the window is closing? Because of the advancements now, AI augmented, if you will. So they're just exponential in terms of the capabilities. And we're talking about open claw, which again, I think is exciting. I don't want to poo poo it, but there are also some real concerns. And so what's next will be augmented by open claw agents building other like the next thing. And so making sure that we're capable of working on all of these things to make AI powerful for people with the kind of guardrails and the alignment that we've talked about. This is the time to do that, but there will come a time when it, if we don't get it right, where it may be too late. Crypto has taught us immutability has costs, no charge backs. Traditional banking has fraud protection and appeals. Where should we be thinking about agent commerce in this? Yeah. So I think one of the big hurdles that people have now, even with agent commerce, just the sort of consumer shopping behaviors is if a merchant tricks my agent into buying something, who's liable? The merchant wants to know this, the consumer wants to know this. And these are questions that are answered in say the visa world or the master card world. And they're not answered yet in agent commerce. So it's very difficult to establish liability paths. And these are, this comes back to me to agent identity and understanding who it is, what are the rules it has to follow. And then who is ultimately responsible if an audit shows that an agent bypassed its guardrails. And so these are things that have to be defined. There are lots of different opinions. I'd say some are very traditional, which is more along the lines of AI is just a surface on the underlying rails. We believe new rails need to be created and that smart contracts and blockchain technology help us get a step forward in creating this. So you said earlier that most economic transactions in the future will be agent to agent. And you said this is something we'll want. We'll want these autonomous agents doing that. Can you, can you make the affirmative case? Like, why do we really want that for all the people who are listening being like, that sounds insane. I think it's not even the most. I think it's all, I think it's every single transaction will be executed by AI in the future. And I think we'll want that to be the case because it will be the safest way to execute transactions. It'll be safer than trying to trust a web form and a purchase through the whole chain and it is today. And that's not the case today. We have a ways to get there, but I think that ultimately will be the case. It'll also be the fastest way to execute commerce at the lowest cost. I mean, we, thanks to things like stable coins, we can move a trillion dollars around the world for fractions of a penny and nearly instantly, you know, connecting those kinds of flows effectively does require a similar level of intelligence because now the bottlenecks in terms of speed are not the machine. It's not the blockchain. It's all the layers on top of it. And then when it comes to earning a return on our assets, I think it will be the only competitive way to earn a return. I mean, today, you know, things like private banking or certain opportunities are accessible to very few people. And the promise of AI is that you have the same level of access that everyone else, all of these businesses, these sort of high end businesses or ultra high net worth individuals have. And the only way that you get there is by trusting these agents and they will be the only way that you get a competitive return on your underlying assets. And so that's this, that's the optimistic vision, certainly the thing that we're marching toward. So one of the things that, you know, you know, I know, technologists know is that these AI systems are inherently these interesting probabilistic systems. But when you get to, for example, you know, kind of a statement of like all commerce transactions flowing through agents because it's higher reliability, how do you square the probabilistic nature of these things where, you know, it's like, well, we've gotten them highly reliable in a certain set of circumstances, but they can go strange very quickly. Yes. With the necessity of reliability within the financial system. So what's your thinking about like what is already present or what needs to be created? What's the thought on that impedance? Yeah. So I think I started the user experience level, which is I think a lot of the tasks that we'll be executing will involve some level of orchestrating these agents doing a variety of things for us and steering and tuning. And then past the user experience is the underlying infrastructure to take that steering. I would say that's more at the AI layer, not just at large language models, but with all of the integrations and data that they will touch. And then when it comes to deterministic outcomes, when they're really doing specific things in the world, you know, moving a car or executing a financial transaction, then those things need to be absolutely deterministic and not probabilistic. And so that's the sort of second layer. But I think that the really interesting thing is that rather than clicking a button in an interface and then checking the readout, the button led to what we expected to happen and ultimately committing, it will be more like we're doing less on those surfaces, but we're doing a lot of steering as needed to orchestrate the agents in the way that we want them to behave, which is what happens in software engineering today. Software engineering is kind of gravitated toward you don't really use IDEs anymore or like look at source code so much because you have a new level of interface, which is, which still requires a great deal of attention. Some would say it's even requires more right now, but you're kind of multitasking across a set of orchestrating surfaces, either on a command line or otherwise to build software. Going up a level, what are the things that we should be doing with our kind of governmental financial infrastructure to kind of bring all of our companies, industries, inventions, society, etc. Ford, what are some of like, if you just said, Hey guys, if you started paying attention, at least a few things, do these. Yeah. I'll say on the policy side, there's a policy component to it. Obviously we're leaning into being a regulated financial institution. So we're confronting the existing policy regime, but when it came to building US dollar coin and stable coins, our belief was then that this is something in order for the for the dollar to work on the internet, the United States government has to say that this is how dollars can work on the internet. And there needs to be a public private sector cooperation and sort of joint innovation. And the way that we usually do that in the West is that there's public policy and there's private sector innovation. We thought it might take 10 years. It ended up taking about seven years to get the Genius Act passed for stable coins, where the government says you have a dollar on blockchain rails. It needs to be backed by these T bills and cash and sort of overnight repos and so on. It's like very specific thing. Here are the consumer protections. Here's how, you know, can't return yield because it's not a mutual fund. You know, those sorts of things. And when it comes to regulation of AI and finance, I do think that ultimately we'll have policy requirements around any AI actors that are permitted to touch things like dollars. And that's where public private sector again needs to needs to cooperate because the US dollar is backed by the United States government. It's not a private sector technology. And so how these things move through agents, similarly to how they move through blockchains will have to be encoded ultimately in public policy. The way that that is happening today is there are several different sort of work streams in DC and otherwise to explore this. Some come out of the AI safety world, AI safety guardrails in general and sort of clearing evils and they're like, others are very much out of finance. And so there are many different, I would categorize my experience so far as they're fairly nascent. It's an important investment toward a long term result that we'll need to have in place. So if we go back to March, 2023 and Silicon Valley Bank, the whole point of a stable coin is that it's stable and that it is pegged in this case, the US dollar. US DC briefly depagged. Can you take us through like what broke, what held, like what did this teach us about stable coins in that moment? So in terms of the stability of a stable coin, we did not have Genius Act in place then. So there are many different approaches to how you provide stability. There was no clear sort of reserve. So you have this, you have this token on chains, but it's backed by reserves. There's no clear guidelines from the government as to how those reserves should be structured. And when we were building US DC over the years, we've had different approaches to how we manage our underlying reserves as did, as has everyone else. Since then, even before Genius Act, we've moved into, into this mode is very conservative of underlying, as long as you have trust in the US dollar and the US government, it's a very conservative reserve backing, which is how we're able to ensure stability. And, but it is the case that stable coins, this is sort of the next problem with stable coins. It is the case that they trade against one another in the marketplace. So we first lacked regulatory clarity to say, this is how we achieve stability and this is how the government will enforce protections for everyone. Now the problem is if you launch a stable coin, you launch a stable coin, I launch a stable coin, we can do that because we have regulatory guidelines, but none of our stable coins will trade as a dollar. Yeah, we have this, it's this sort of a singleness of money problem. If Amazon issues a coin and Walmart issues a coin, well, they're, they're, they're going to have a market against one another until we solve this interoperability and this sort of clearing house issue. And that hasn't been clarified by the Genius Act or anything else. That's what I was going to say. So in order to clarify that, do you think we need additional government regulation? There's some technical approaches to it. So I think we'll see some solutions that people can rely on it by people. I really mean those businesses. You just want to look at this as a dollar. They don't want USDC, but they want as a dollar that moves more efficiently and is more accessible in their markets. And if they have 20 of those, they want to view all of those collectively as just dollars. And so there are some technical approaches to how to, how to achieve that. But we may also see it addressed in subsequent legislation that comes out in this space. So many crypto maximists, you know, kind of worry that circle other similarly centralized stablecoins can lead to a government having full control over individual citizens ability to transact, monitor all transactions. Do you see this as a risk of heading to a CBDC surveillance world? Is there anything we should be doing to navigate in or out of that? So on CBDC, just to clarify, central bank digital currency, there's this idea that maybe the federal government should issue its own stablecoin and issue it directly to individuals. And that's been an approach that has been looked at, say, in China. Here, generally, the way we do things is that we have private sector innovation, but with public policy. So we have the government saying how to do it. And then we have companies such as Circle and others that are able to actually do that thing. So that's the kind of difference between stablecoin issuers in the US versus the sort of CBDC approach. But the crux of the question really, I think, is related to privacy. And it is important to have an open mechanism for value exchange that also has privacy protections. And a simple example of that is if you you're paid on chain, I shouldn't be able to see your paycheck. And so there needs to be some privacy protections. Another example is for very large transfers internationally, there's counterparty info exchange, or there's this thing called the travel rule, even at relatively low volumes. And so that should be private. You don't need to know my address or social security number if I'm part of these transactions. And so there isn't a there isn't a pair to maintain privacy and balance transparency on these open rails with privacy. And so this leads to things like Circle has created a new blockchain called ARC, which is really designed specifically just for stablecoin payments, as opposed to all the other things you can do on other blockchains. And there have been others that have looked at the same space, Striped Bridge of Form Tempo, and we also leverage that on Testnet. So, you know, as Catayna, we're we're we just want a solution in place. We're not necessarily weighted to one, but but part of the issue that these payment chains are addressing really is this privacy issue to avoid the surveillance state on chain. So you talked about how you thought it would take, you know, sort of 10 years to get the legislation that you needed, and it took, you know, seven or eight to get the genius act. And it was it's so critical actually to understand the rules of the road and to have this public policy so that we can have innovation, which I think is is sometimes the people don't think about that. They think about innovation and regulation actually being at odds, but in some places in order to have innovation, we need regulation, especially when we're talking about people's finances. And I think a lot of people sort of take the financial regulation for granted because it's just what we have. And if there's fraud on our Visa card, you know, they'll, you know, refund us, you know, sort of all those things. Where do you see this landscape for Catayna? Like how long is it going to take? Like what do we need to have happen so that we can have the regulation that causes the innovation? Yeah, I'll say how we're going about it because I do think it's similar to the circle, but it's different from other tech companies. So you typically the way you build things at tech companies is there's some version of sales and marketing, BD partnerships, that sort of function. There's some product manager, product design, and you know, and then there's the engineers and the three of them are the big stakeholders and everybody else is kind of supporting. And when you're building something in finance, you have a fourth stakeholder at the table, which is related to risk and compliance and in these elements. And things go wrong when any one of those stakeholders is kind of out ahead of the other, you know, sales is selling things that can't be built in a reasonable amount of time or product managers create, you know, documents and they have a cabal of decision, but it's not, you know, our engineers build amazing tech. Nobody wants, you know, that kind of stuff. And then the same can happen with risk and compliance. So he needs to seat at the table and it needs to be moving in lock step. With the other, the other three. And so that's, it's kind of a compliance versus how people refer to it. But, you know, compliance first sort of approach to building out this technology is the same one that we're taking at Catena. Taking it at Circle certainly causes us to be slower on certain things relative to others who are in the crypto space. But we were absolutely had conviction that it was the only way ultimately to have the winning solution. You've built two generational companies at two massive inflection points. You could have stayed at Circle and worked there a long time. You're still on the board, which is fantastic. But what drives you towards the chaos of a startup and starting something new? So I give the healthy and unhealthy answer. I'll give the unhealthy answer first, which is just, it's almost like an obsessive disorder. I can't not do this as I tried not to do this, but I have to do this. You know, I build things. I'm not, I haven't been that successful investing in other people build things, which is a great skill I admire in people, but I build things. And so I have to build this. And then the, the healthier answer is, um, I do think again, that we have a window to get this right. And I feel like I can play at least a small part in getting it right. Um, and that's a combination of, of understanding the policy, the underlying tech, how to bring the right people in and, you know, sort of having pattern recognition for what to apply to this stage. And it's just, I'm excited to get out of bed every day and work on this problem, uh, in this space. Absolutely. Awesome. So I'm rapid fire. Is there a book, movie or idea that gives you optimism about the future? Usually I answer with books, but I'll give a movie answer because that's what popped in my head. So there's a movie called Arrival. It was a tiny, uh, Ted Chang short story. Yeah, exactly. Yes. So, so it was kind of like an alien first contact movie, but the thing that I think would give, it gives me optimism is, is, is built on this. There's a language that they need to understand and the language as they begin to understand that doesn't just allow them to communicate. It changes the way they think, um, and opens up new patterns. And, and so, you know, the thing about developing stable coins is it unlocks new kinds of business opportunities that can be built with programmable money. AI is like, we're just learning the grammar today to how to understand this language. And it will unlock new ways of thinking, uh, for all of us. So I'm optimistic about that. Awesome. Uh, what is a question about crypto or AI that you wish people would ask you more often? I guess both. And it's kind of a personal answer, but, um, you know, my, I have a background in my undergrad degrees in English and, and I was, I'm a musician. So, you know, and so when usually when people hear that they're like, oh, that's interesting, but I can see what they think of, like, maybe he didn't know what he wanted to do or there was career drift or these sorts of things. And so the question I wish people would ask is how's that connected to all the products and companies that you've built. And because I think it's very directly relevant, especially in the age of AI, when sort of, you know, skills across multiple domains are incredibly valuable as opposed to just sort of like climbing a career ladder, which is, you know, those two, I think we'll find their tasks most in jeopardy. Right. Instead of leaving your field to start in this field, you use that knowledge, which was incredibly important to what you're doing today. Absolutely. The music thing is fascinating. Cause I've been thinking about music a bunch. Where do you think the current state of AI music is? What are some of the things that are kind of perhaps off the beaten path of, Oh, is it an issue with the industry and the creatives issues that navigate? Is there issue with copyrights or training or inspiration in order to be generative? Anything that that has kind of struck your musical background? I think a lot of the things that are circulating are really good replicas of other things. That's not entirely bad, but it's also kind of empty. So, you know, it's like lyrics that are a little sloppy, but it kind of bops, you know, it's we're kind of, there's a lot of that going around, which is a little more like an echo of originality and that may come from a number of different originators who have obviously are not being compensated and there's no mapping back to that. There's also experiments more along the lines of sort of electronic music and experimentation where it is a skill to compose with sounds that you yourself are not capable of making as an instrumentalist and there's still a creative element to it. And so I would hope that the future of AI music moves more in that direction as opposed to just, you know, the millionth copy of Let It Be sung by Cardi B or something, you know, whatever it is. Exactly. So where do you see real progress happening outside of your industry? At the side of my sort of financial and even AI foundational domain, an industry I'm really hopeful about is related to health care and sort of the cost of producing drugs coming down, accessibility to health services, you know, sort of increasing. It's not a domain I'm deeply familiar with. I have friends who are deeply in that space, but it's also it's one that I have personal connection to. I lost my sister from cancer, I lost my mother from cancer and sort of just that personal experience of seeing how the databases are just not connected and nobody has all the information that they can even use to guide people. There are trials that are available that people don't even know exists and they can't, you know, sort of connect. So I'm very optimistic about some of these technologies being able to clear those gaps and it ultimately comes down to accessibility, which is the same thing that sort of drives a lot of my interest in finance. So as always, our last question, if everything breaks humanity's way in the next 15 years, what do you think is possible to achieve and what's the first step to get there? Wow, 15 years. So I think I'll steer it back to what we're working on. So if we get this right, then I do think there will be kind of a hyper-personalized private banking experience that will be accessible to businesses everywhere, whether you're an entrepreneur in Nairobi, you're a two-person company in Detroit, or, you know, whatever it may be, you have the same access to the global financial system with new opportunities to start new businesses, tap into, you know, new mechanisms to go to market that you just haven't had available before, thanks to the combination of AI agents and underlying financial infrastructure that can make them safe to transact. Amazing. Awesome. Great talking to you. Great talking to you. Thank you. I enjoyed it. That's fun. Possible is produced by Palette Media. It's hosted by Ari Finger and me, Reed Hoffman. Our showrunner is Sean Young. Possible is produced by Tenasi D'Elos, Katie Sanders, Spencer Strasmoore, Imou Zou, Trent Barboza, and Tafadzwa Nima Rundway. Special thanks to Surya Yalamanchili, Sayida Sapieva, Ian Alice, Greg Beato, Parth Patil, and Ben Rallis. And a big thanks to Victoria Lampson and the Lighthouse in Venice.