When Bots Have Bank Accounts: The Rise of the Agent Economy (With Sean Neville, Catena Labs)

You co-founded Circle and now you're working on something new, an AI bank. The internet self is going agent native. Really agents doing anything. We think they'll need to get paid. They'll need to make payments. They'll need to generate some sort of return or be rewarded for holding a balance. They may want to lend. They may want to apply for credit in certain situations. They'll want to do all the things that a bank might do for a business. Ultimately, we'll need an AI bank. It is actually for other AI's. You believe that these agents are going to be the only ones that we actually trust with our assets. What is going to make people one over by bots? Why would you ever trust it with your money? Worst outcome is you build a lot of amazing stuff that nobody wants. Sean, thank you so much for being here. This pleasure. You co-founded Circle, you architected USDC, Debalcoin that lots and lots of people throughout crypto and beyond are now using as crypto and tradfie sort of intersect. And now you're working on something new. You're building what you've described as an AI bank. What do you mean by that? Well, we're taking the next step. Once we have stable coins and we have the ability to represent dollars on internet rails, what kind of new opportunities are unlocked? At the same time, we were contemplating that you could see clearly the web, the internet itself is going agent native. And after working an AI for quite some time, develop a conviction that as AI actors become economic participants, they will ultimately be a primary dominant economic participants in the world for all kinds of activities, payments and otherwise. I think flash forward in a few years, I think they may be the only actors that we trust with our assets and the only actors that are capable of generating meaningful return on our assets. We're certainly not there today. And so what do we need to do in order to get there to unlock, I think, a level of prosperity like so which the world hasn't seen yet? And as far as out, there's some fundamental things missing. So one of the things we're doing is working on some infrastructure to make it safe and trustworthy for AI actors to participate in the economy. And then we're building on top of that foundation, what has been called an AI bank or an AI native bank. So that's what we're working on. So you said that you believe that these agents are going to be the only ones that we actually trust with our assets. I mean, we're living in a world right now where there's famous gallop poles about trust in institutions just crashing, utterly crashing over like recent decades. Trust is a hard to come by resource nowadays. It was similar in some ways to trust in money flows. You know, the old way of managing trust was, let's have a whole bunch of regulations that tie humans in the businesses they create to a set of rules so that at least when they prove that they're not trustworthy, we have sort of clear liability paths and repercussions. And now we have something that's improved an improvement on that, which is we have the ability to encode trust into software using cryptography on rails that no one can control. And it's a common good, a public utility for the world. And so, you know, similarly, when we look for to semi-autonomous or autonomous actors that are acting on our behalf to do all kinds of things, a hyper-personalized, you know, bank for you, there's different from me, but those things interact. What are the elements of trust to enable that sort of thing to happen? Because you mentioned the word bots. Today, one of the examples where that world is impossible is if you look at the existing risk infrastructure in financial institutions, which is designed to make sure no bots can participate because they're all bad. Make sure you're a KYC individual or a KYB business. I use the word intentionally because people have a bad association with it. Absolutely. And for good reason, but what we really need is a system of risk that can say, let's assume the only participants will be bots, but still keep the bad bots out, the bad actors out, but have some way of identifying the good bots that we want to interact with. And then, beyond that, apply policies to them. So we can say, you know, I would like to interact with the Amazon bot agent. No one can agree on what an agent is, but you know, we'll stick with the word bot for now. The first time I had people were talking about the agentic, the kind of, it was like, what are you talking about? Well, agentics sounds like agent-ish because no one wanted to, you know, there's a lot of infighting around how do you define an agent? So, I was like, was this like James Bond stuff? Where are these agents coming from? What are we talking about here? Well, I can give you how I would define an agent, but everyone seems to, everyone in the AI, you know, domain seems to have a slightly different view. But I do think when it comes to things like risk infrastructure, we need a way to verify agent identities in a way that says, who are I want my agents to interact with? And how can I prove that it's actually, how can I prove that it's actually Amazon's agent that I'm interacting with? And then how do I give it policies to say, you know, spend up to $200 in this time period, but don't spend more of that asking me, like, how do I do those things? And please don't make any paper clips. Yeah, and then like, and just, and help me understand why you're making these decisions, so a degree of auditability after the fact. And so, these things don't really exist yet at the foundational layer of agent activity. We're still struggling with, you know, people's anecdotal experiences with AI is they use, I'm not going to pick on a particular one, they use a general chatbot and it hallucinates a terrible recipe for chocolate chip cookies. At the same time, they hallucinate more money in my bank account. It might be more open to it. The integration is actually, I think, one of the big problems now. The LLIMs, the foundational elements are good enough now to make an level of decisions that can be judged and sort of evaluated, but the integration with the real world executing APIs or tools and applying policies to those is is much more difficult. You can, you can sort of demonstrate it, say an agent, even a chatbot, executing a particular tool, buying groceries or whatever it is in a demo that is impressive, but it kind of obfuscates the fact that making it truly reliable is extremely difficult right now. And there is a way to do it and we're getting better and better at it, and this is the worst it will ever be. It will only get better, but that really is when it comes to trust, those are the kinds of issues to resolve with really agents doing anything, agents in healthcare, agents managing air traffic, you know, whatever, but particularly in finance. Wow, air traffic, yeah. You mentioned earlier the sort of spectrum of going from semi-autonomous to autonomous. I feel that it lies a lot of, you know, what needs to happen along the way. What's the state of the technology now? And how do you expect it to play out to get from training wheels to full on, you know, autonomy? So I would maybe couch it in a couple of ways. One is related to what we're building. Now, I could say in a top-the-foundation. You know, our belief is that ultimately workflows that have agents as participants. So groups of agents can post together to do things. We think they'll need to get paid, they'll need to make payments. Once they're doing that, they'll have access to funds, they'll probably need to manage effects, whether it's between even different stablecoins or effects to stables or traditional effects. They'll need to generate some sort of return or be rewarded for holding a balance and interesting. So they need to do all of these things. They may want to lend, they may want to apply for credit in certain situations. And so ultimately, we'll need an AI bank. We're not there today. But that kind of service imagines, you know, something that is a much more autonomous system that can execute financial activities safely. Still with humans in the loop, you know, acting autonomous doesn't mean that there aren't checks with humans for certain things. Simi-autonomous, the line begins to blur between humans in the loop and AI in the loop. Simi-autonomous or activities where LOMs may be planning, executing a plan, making a plan and then beginning to execute on it using tools, but with heavy human interaction. So not consistently running that loop forever. So there's a little bit of a spectrum of how you define humans in the loop with workflows that involve LOMs versus workflows that involve maybe some humans in the loop and that's a little bit of thing. So that's generally how we think about it. Where we are, I think in certain domains, we've seen a lot of progress with AI as economic participants. But what we haven't yet seen at scale as we're talking today is agents actually paying agents. We're not even really seeing agents communicating with other agents yet. We're seeing agents pay for access to APIs or access for data or paying humans. We're seeing humans pay into AI workflows in some cases. But we're not really seeing agents, these compositions of agents working together as Simi-autonomous or LOMs entities. And so there are some things missing to even make that possible. There's a lot of discussion around how do agents even communicate with each other? If we're just interacting with a chatbot, we're using a web browser and it's a human interaction with some sort of HTTP interface. How do two agents talk? Is it going to be over HTTP? But what if they need to talk over SMS or over voice lines or various other mechanisms? We haven't defined those things yet. We basically haven't defined the SSL for secure agent agent communication, let alone payment flows. And so we still need to address that foundational layer. And there's a lot of debate about what that layer looks like. What I ideally want to see standard that we can all say, hey, this is how you do these things. Few and I go build a e-commerce shop. We put it online. We don't really have to think about SSL or do we have to do it the Microsoft way or the Apple way or whatever. It's just a common standard. We can all go implement those things. Right now that kind of thing just doesn't exist. It'll offer it for agent development. There are a number of people putting forward standards, ideas for early standards and protocols like X402, MCP. What is your sense of how this is going to shake out? I'd say it's more fragmented now than it was three, six months ago. When it comes to things like representing agent identity or KYA or know your agent as we're saying and a lot of different approaches to payments. Some are approaches that have been promoted by I would say large incumbents, but in a way that doesn't necessarily bring other incumbents on board. And some are by groups that are more I would say on the start up into the spectrum. I think there is interesting collaborating on some of these foundational elements because it stands to grow the value pie for everybody. But right now in AI it's so things are moving so quickly and it's so early that I think it's unclear exactly where where the value will accrue. And so a lot of groups are trying to own the entire stack because they don't want to necessarily give away the place where the value may accrue, which is not unlike where we were back in the days of online services before the internet exploded. And there was America online and prodigy or whatever. And we could send email to each other if we were both on AOL, but that was the only way. And we paid 15 bucks a month or whatever it was for the privilege. X-Work 2, you mentioned, is a payment protocol that is interesting. There have been people who've kind of come on board with versions of it, but even the different versions are a little bit fragmented themselves. And some of the larger players have not yet come aboard. So for instance, there's a lot of focus on X-Work 2 for allowing scrapers, either for pre-training or even at inference time, to access data as a potential solution for paying for content acquisition, which is a great concept. But if no major publishers come on board and no major scrapers come on board, then it's just tech. And so the hard problem with solidifying some of these foundational standards is not so much the technology. It's not, is X-402 better than this other protocol. It's more who can bring together at the table, the right parties who normally really hate talking to each other to agree on an approach. - What lessons have you learned from building circle that you're now applying to your new company? - Well, one lesson that I think we embraced at circle is relevant to developing AI. I was sort of framing it in that sort of context of delivering a product. And so at a typical software company, there are usually like three dominant forces at the table and then a bunch of supporting forces. But the three dominant forces are engineers that have people write the thing, product people, which is your background. My background is engineering, yeah, I was software engineer. And then product people, and then some form of sales from marketing, partnership development. Usually those three are the key sort of forces and you really want them all to work in tandem. In crypto, I think, it's often the case that engineering is ahead of the other two groups. And the worst outcome is you build a lot of amazing stuff that nobody wants and doesn't end up having a lot of utility value to other people. The danger of sales and marketing being out ahead is you build things that everybody wants, but they can't actually be built, I mean, at least not in any reasonable timeframe. And so you're sort of selling vaporware. And then the product one has a few issues, but if the product team is sort of out ahead of everyone else, you end up with this kind of elite cabal of people who create reports and slide decks and they go to the conferences and they talk about all these things. And then you'd really never hear anything more about it. It's like it's in Discord or it's tweeted. There's no substance that ever really emerges. And so really you want all those three together. The lesson that I was going to go to a circle is there's actually a fourth major stakeholder at the table, which is the regulatory component. And that's very different from other kinds of software. And they're not just supporting. They're not G&A in the background. Actually are a major stakeholder supplying requirements and helping facilitate development of a product that can do great things. Still important post genius. Even more important because now there's a regulatory playbook that people can march toward. It was much harder before, certainly in the stablecoin space, when institutions, customers, businesses were just unsure how the United States government viewed US dollars on blockchains. Are they securities or the funds? How, what are the consumer protections? How do the reserves need to be configured and all those things are prescribed now? So it's much clearer. Just so we can actually come to the table and say, here's the playbook. We actually have one that will get the US government stamp of approval so that anyone who is accepting a payment dollar stablecoin can rest assured it as a payment dollar stablecoin that the US government stands behind. I was going to describe what we were saying earlier about engineering product and sales and marketing. It sounded like a three body problem to me. You've got these three different entities that their own gravitational attractions and trying to manipulate them so they're working in tandem and understanding how they function together. Yeah, I mean, that is the puzzle of building a successful software business, is balancing those three. But it is the case, not just in finance, but in other industries, healthcare, and other heavily regulated industries that the regulatory piece is an important strategic element to think about when you're doing a product roadmap, when you're actually executing. I've got to ask you this, as somebody who invented one of the most important stablecoins out there right now in most significant stablecoins, you don't like the term stablecoin. Why? I've never loved the word stablecoin and I think that maybe at some point we just stop using it. I think a lot of people who are in the non-tripped or capital markets use cases who are attracted to stablecoins really want dollars. Now, of course, there are other stablecoins beyond dollars. There's eurocoins, there's wine, there's peso. But most of the demand right now is not really for a stablecoin or for USDC or any of the others. It's for dollars that run on internet rails and run at internet speed that are fully regulated and programmable. And so it's not really a stablecoin, which is a reference to stability relative to crypto. The price is meant to be stable and for the dollars, back to the dollar, it's really more a reference to dollars on the internet and we don't have a smarter, sharper, like really pithy term for that 'cause stablecoin is easy to say, it kind of rolls off the tongue. But that's kind of my little pet peeve around stablecoins. On the AI community, you know, as we were working with AI engineers, before I would say the last six or seven months, we avoided using the word stablecoin because it was so heavily associated with crypto, which still had some negative connotations, I would say, in the worlds that we were working in. And that has definitely changed. And they've actually, now I think people generally are embracing the word stablecoin. So maybe that fight is lost at this point. One of my favorite points on this is there's a story from a big national media entity and the headline was, there's a new cryptocurrency called stablecoin, here's what it's all about. - Yeah. Well, so we put together a USEC in 2017, 2018, rolled it out. And then it wasn't really a success in terms of market cap and transaction volume on say until 2020-ish. So it took a little time, but that's still five years ago. So, but, you know, better relate to the party than never arriving, I guess. - Adoption has picked up since then. - It has. - You have time for a quick lightning round? - Oh, let's try it. - Okay. What's the worst advice you've ever received as a founder? - So the worst advice, it's actually good advice, but it's not great advice for what we were trying to do particularly with circle. And I think the same with Ketana, which is find the one customer that has a specific pain point that you can address. And the reason it's not great advice, it is a good advice for certain kinds of projects, but the reason it's not great advice is because there's a tendency to over index on what a friend of mine calls like SAM from Accounting. And you're solving SAM from Accounting's one problem. And then nine to 12 months later, you're still solving SAM from Accounting's problems and building for SAM for Accounting, which might get you to a million ARR or something. But if you're trying to build a $10, 30, $50 billion platform company that's doing something generational meaningful, you can think of multiple customer types at once. And you can think, you know, you have platform think about it. So it's not actually bad advice, but in our case with what we were trying to do, it would have led us to, I think, oversimplify on some experiments that were very meaningful in getting us to where we are now. >> It's funny 'cause we were talking on stage with Zach Abrams of Bridge earlier today. One thing from his story is they landed their first customer, this like Colombian company, and it dictated so much of what they ended up building and actually, you know, enabled them to kind of get where they were today. It's funny 'cause it runs kind of counter to that stuff. >> Oh yeah, I think it's very common advice to find that pain point. Be a pain killer and out of vitamin, that's the cliche. And it's a cliche for a reason, because that's the way to build companies, but for me, and for what we're trying to do now, and I think, you know, for Jeremy and for me at Circle, we were thinking a little bit differently, and we've been criticized still today, we've been criticized, you guys were trying to do too many things. It's very deliberate in what we were doing, and it absolutely led to, you know, what we've been able to achieve so far with Circle. >> Given that you're working so much in AI, I'd love to ask, what is your biggest productivity hack? >> So I use all of all the models. My favorite is a local pipeline that connects many of them, and this is a total hack. This is not something that's that different from what a lot of people in AI and senior circles do. The local pipeline, I start with Cloud, so the Anthropic models. And then I have a series of Gemini, and OpenAI models judging the output from Cloud, so constantly sort of think of them as, you know, the teachers slapping the knuckles of Cloud to try to do better, and I put, I workshop through it, and I ask it, you know, personal things, sort of work through different pipeline, different memory contacts, sharing projects, and the like, and so it's not one model though. It's not, you use ChatGPT for this, and then Cloud for this is use them all with different contacts. >> So you constructed like a Rube Goldberg LLM, so Rube Goldberg is sort of a local mouse trap. >> Very nice. >> Yeah. >> Give us a book recommendation. >> I would say my class would go to for start of people in particular, is been start of things about hard things. >> So good. >> Yeah. >> Classical. >> It's great. >> He mentioned advice that he thought was, you know, really good answer to the question of what's bad founder advice, which is higher A people, like, this is great advice, but I mean, I wasn't gonna hire a bunch of more. >> Yeah, yeah, thanks. >> Yeah, so that's always, that's always what I recommend. >> You know, rewrite on. >> I mean, obviously, you, it sounds a little self-serving, but genuinely that's the book that I pass out to people who need to learn a little bit more about the reasons that what we're doing, Ecrypto was so important to the world. That's another great one. >> That heartens me to hear. Thank you. I hope you're not just saying that. >> I'm not absolutely. >> Finally, last question. What is the smallest hill that you will die on? >> The greatest food ever invented is sushi. >> Okay. >> I think I might agree with you actually on that. I might die on that hill too. >> Yeah. >> What is your favorite of the sushi's? >> So I'll eat anything. I, the closest I would not eat. >> I would not eat the funnel of sushi. >> Your sushi with sushi. >> Sushi eye. >> I have no idea. >> The closest thing I was on the borderline about was seeing them any. >> Seeing them any. >> Wait, is that, that's not the one that like can kill you? - No, I think that's like the poison bladder in it or your stomach. - Pufferfish or the other part of your stomach? - No, but it's a little spongy. It's a little like a moose sushi of like a pate. - Yeah. - As a closest I've ever come to not like, so I think it's the universe, it's universally the best. - You'll kind of rubs me the wrong way too. I'm sort of like, I'm going to go with you. - I don't know if it's like kind of like a snake or something. - I'm good with it. - Yeah. (laughs) - Thank you so much for your time. - That was a great pleasure. - Yeah, thank you. (upbeat music) (upbeat music)