From Building USDC to Agentic Commerce: Sean Neville on Programmable Finance

Cuy Sheffield  0:00  

Kai, welcome to Tokenized, the show focused on stable coins and the institutional adoption of real-world assets. My name is Kai Sheffield, I'm the head of crypto at Visa. Welcome to another episode of our agentic commerce series. Back again, stepping into the co-host seat, the CEO and founder of Mesh, Bam Azizi. Good to see you, Bam. How have you been? Thank you, Kai. I think it has been a while, but excited for this episode with Sean. Sean, thank you for taking the time.

Sean Neville  0:37  

Yeah, my pleasure.

Cuy Sheffield  0:39  

Joining us today is Sean Neville, CEO and co-founder of Kataina. So, two quick bits before we get into content. I need to remind you that the views or opinions of our contributors today are their own and do not necessarily reflect those companies they represent. Nothing we say should be taken as tax, financial investment, legal advice. Do your own research. And a reminder, this podcast series is made possible by Mesh and Visa, all right. We've got an absolute legend in Sean here today. I learn a ton every time we talk and get to Rif. So, why don't we just jump into it? Tell the audience a little bit more about your background, some of the founding story of Circle and USDC before we get into what you're doing today.

Sean Neville  1:18  

Sure, yeah, so my background is in software engineering, and prior to Circle, I built a bunch of things. When you know we were starting our first companies, back in Circle, is actually the third company that my co-founder, Jeremy, and I have been associated with together. In our early days, we were building out, you know, early companies in the early days of the web, which I know ages me a bit, but and then subsequently built some, you know, products for companies like Macromedia, Adobe protocols, languages, mobile apps, but really gravitated toward what can we do to make finance run more like everything else on the internet, and that's what ultimately led to Circle. We founded Circle in 2013 We didn't create USCC in 2013 probably should have been more obvious in retrospect, but it took us four years to realize that we needed to put dollars on blockchains, and so did that in 2017 and yeah, I guess the company is it's 13 years in, which is like 130 years in crypto years or something like that, but it's been, you know, quite a journey, and then obviously I've gotten very excited about what's next in finance with AI, which is what we're focused on at Katana.

Cuy Sheffield  2:22  

Before we get to Katina, take us back to 2017 and I believe you wrote the first USDC smart contract. I think you were telling me you wrote it on a plane, or there was there some story of, like, you actually created what became USDC. How did things differ from what you thought at the time? You know, looking back when you were launching this into the world, like, what were some of the surprises in ways that USDC grew and played out that were different than you expected?

Sean Neville  2:48  

Well, when we started Circle, even before USCC, I think we had conviction around broadening access to the global financial system, and that money is just data now, it should be able to run on open rails, and, and then specifically with USDC, the form that money would take would be programmable dollars on these underlying rails, but we thought both in 2013 we founded the company, and we started working on USDC, we thought for sure the first use case that would be tackled would be payments, and it obviously is happening, but it's not happening at scale even yet, it still feels like we're fairly early days, so I think we were right about the overall vision, but in terms of how the vision played out and what the sequence is, we weren't right about everything. The dominant use case for stable coins has still been crypto capital markets, so that was one thing. Also, I think for USDC specifically, it was important to us to not create like a vendor coin, like a circle dollar, but that this would be approached more like an open standard that multiple companies could build on, and that meant founding a consortium. You may remember, of course, you remember you were involved in a lot of those discussions, and the idea was that let's just agree on a standard way of putting dollars on the internet, and then we can all compete with each other in the market at a higher level of value and grow the pie for not competing at just the foundational level and there shouldn't be 20 different versions of dollars on internets that all trade against each other but there should just be dollars on it on the internet that we can all build products and businesses with and so that you know that was the take I still very much believe in open standards and approaching these things through that lens, and then the other thing, of course, is if it's $1 on internet rails, and that means that ultimately the United States government needs to weigh in on how they define $1 running on these blockchain rails, and so there was a policy initiative that started very early on, thought it might take 10 years to sort of enshrine in public policy what $1 running on blockchain rails would be. It took about seven years with Genius Act last year. So, actually, I had a schedule, but I think we did, to our credit, realize early on that there was a policy piece of this that was critical to invest in. But you have to remember, too, at the time, the business model for the stable coin was a little questionable, because interest rates were not so. For great, you know, as we thought, maybe it takes 17 years for those to turn around, and so it was a lot of conviction early on when we were writing, you know, the first white paper, the first code, kind of thinking through right partnerships and how distribution would work at a time when there wasn't really clear public policy, huge regulatory hurdles to invest in clearing, and then also the fundamental business model associated with it, and so all those things were challenges at the time. I don't know if you say, "Hey, we were right about this and wrong about this. You can if you go into the details, but ultimately I think we're pretty happy with the trajectory that it's taken.

Cuy Sheffield  5:37  

Yeah, it's been amazing to see the success of the product, and I feel like people forget that Genius Act happens much faster than most people in the industry expected. Like, if you would have told me three years ago that we would have a federal law around stablecoins, like it's a dream for many people who were building early on. What was the founding story of Catena then in kind of your journey post circle? What was the initial problem you're trying to solve, and how did you first just go down the AI and energetic rabbit hole, kind of what pulled you into that space?

Sean Neville  6:06  

Yeah, I mean, it's funny, at that time AI did not yet mean large language models, and so our, we have pretty deep backgrounds in AI, but it was from the ML side, and building out risk infrastructure that is suitable to manage sort of conversion of all the fiat monies into stablecoin monies and other things, and not large language model AI, but we did develop conviction that in the future stable coins made dollars programmable, and AI actors ultimately make finance on the whole programmable, and we could sort of see a world where, if we get it right if we get AI right. Generally speaking, and sort of aligned with our business interests, and we could unlock a new kind of prosperity, though, like, so which we just haven't seen, where most of the economic actors, if not all the economic actors executing the mechanics of transactions, would be some form of automated system. And then, of course, as we started believing this large language model, the whole world sort of embraced LLMs, I wouldn't call them agents necessarily, but workflows with increasing with large language models involved in decision making, and so as that began to happen, it sort of moved from thesis, is it possible to connect an agentic actor or an AI workflow to a bank account or to a wallet or to a card, and now it's sort of progressed, where it's no longer a thesis, you can actually, that turns out to be the easy part, because really connect it, doing it reliably in a governed way that's fully auditable and really controlled across multiple layers, that's that's a hard problem, and that's really the problem that we've zoomed in on with container, is making it possible for businesses, enterprises of all sizes who are building these agentic workflows to move money safely across multiple rails, of course, including stable coins, but not limited to stables.

Cuy Sheffield  7:55  

What's your kind of starting framework for the state of agentic today, and like, how do you separate? I feel like when we've talked before, we spent a bunch of time on like the concept of agentic treasury, and like agentic treasury versus agentic commerce. Like, a lot of people, they think about a shopping assistant buying shoes, it seems like you've been focused on other use cases that you see as kind of more near term or more productive. How do you look at the use cases that agents can then become economic actors around? And then what was the process of narrowing down which of those use cases you, you really start to go after?

Sean Neville  8:33  

Yeah, I mean, I think ultimately AI actors, they'll certainly need to pay, they'll need to pay each other. That's not happening today at scale. They're mostly paying for resources, access to resources, paying through API gateways, and sometimes paying out to humans things like this. Sometimes accepting Ks from humans, they're not really agent to agent commerce yet. This is so early, but I think conviction is they will be doing that. They'll need to get paid for the value that they bring. They'll need to turn a return on the assets that they're managing, they'll need to handle effects, they'll need to handle clearing the sweeps across rails, they'll basically need to do all the things that a business relies on a bank to do for them, and so that's almost the easier picture to see, in a way, they will need to have access, safe access to all of these things, and that does imply, in our view, that things like compliance and risk infrastructure needs to be rethought, not sort of tacking on an agentic API or an MCT server onto a fundamental like ledger and compliance system built for human businesses, because a lot of the agentic traffic looks like fraud or layer in the traditional sense, but really rethinking the fundamentals in order for all of these things to be possible, but then the harder question is, well, what comes first? Like, what's the sequencing? Does this kind of payment say real-time payments for access to services come before you know more broader-based, say AP or legal agents paying human lawyers for reviewing the data says, well, what happens first, and what's the sequencing look like? I think we made the same of. We sort of had the same view in the early days of this that we had with USDC, which it'll be payments, you know, whether it's consumer retail commerce or whether it's B2B, and we were sort of surprised to find that in terms of the automated systems where where LLS are really involved in moving the money and not just sort of participating in the decision making process, actually moving the money, it was not retail commerce, it was more bb, and the immediate use cases that we began to see were not so much payment flows but other things, such as treasury management, sort of liquidity management, risk management, yield, these sorts of things, and so it's similar in a way to what we saw at stable coins, and that stable coins became very, very useful as a way to not just hedge risk, but get out of crypto capital markets, and, but stay in dollars, and so that you could get back into crypto. It's certainly a very valid use case, but you know, the real promise of dollars was something more like global transfer of dollars around the world, you know, borderlessly for practically no money, which is a huge value prop, but that wasn't the immediate wedge use case that we saw. It was more like a risk management instrument. So, similarly with AI, the initial use cases, and I think we just, again, take a very broad, broad view of this, but the initial use cases were more, you know, more in the realm of sort of treasury management. But I've already expanded from there. I do see more, you know, a lot more discussion that could be a little bit of bias based on the lens that we have, but I do see more movement on the business side than on the retail consumer side. I think there are a lot of challenges on the retail consumer side, but the commentary across all of them is really, how do you clear trust, how do you make these things actually trustworthy and reliable, and what happens if they escape their guard rails or their policy constraints, like, you know, what does that, or does that look like somebody underwriting that risk? And these are all the questions, whether on the retail side or on some of the enterprise money movement side, that are the hardest problems to clear, especially for stable coins.

Sean Neville  11:55  

Now that we have Genius, used to be the regulatory piece was really the hardest to clear, and now it's really on the AI side, it's clearing the trust hurdle more than anything else. Stable point still has other issues. Liquidity is a big one. When getting the whole discussion of will there be 2000 stables under Genius, or will they be still the only two that have any meaningful liquidity? That's a whole other thing, but in AI, it's mostly about trust.

Cuy Sheffield  12:19  

Bam, how do you think about this, particularly on the yield optimization side? It feels like crypto and DeFi, and like the ability to have many different liquidity pools and sources of yield. Even before agentic was a buzzword, there were vaults that were enabling you to automatically rebalance and earn it. So, like, are there elements of crypto today that kind of give a view in towards what more institutional treasury management that's automated could look like in the future.

Bam Azizi  12:49  

I'm like, really love the way Sean framed it. They said, like, stablecoin is basically a programmable version of money, but agentic commerce or AI in general is going to make the entire finance ecosystem programmable. The only part of them, like, I'm not 100% agreeing with Sean, is like, I believe, like, in near future, 99% of transactions of agentic commerce happens on blockchain, so maybe it's because, like, I'm very bullish on blockchain technology, but in the interim would be like a hybrid situation that will have like traditional rails and also stable coins for the settlement and payment and whatnot. What I am interested, because my background is identity, unlike Sean, I'm also a computer and data scientist, and my first focus was identity, is still like two days mesh as an identity play from our perspective. What I'm interested to see how it plays out in the agent ecommerce side is the identity side, or as Sean mentioned, the trust aspect of agent ecommerce. We have a traditional term in identity called triple A: authentication, authorization, and accounting. So the way think about it is like the user or an enterprise or business would allow the agent to authenticate on their behalf to have access to their wallet or their bank account, and then the authorization is the level of access, right? So you can spend like $100 or $1,000 and then the accounting is a session, right, for how long you're going to have this authorized access to my account? What I think blockchain can do for Agenda Commerce, beyond the stable coin, is building that decentralized smart contract that enables agents to be able to spend money on top of that smart contract, meaning like they don't get like a blank check, they can spend money as much as they want. I think that part of identity or authorization will happen on chain, which is not related to stablecoin, but that's something that blockchain can offer to the agentic world.

Cuy Sheffield  14:54  

Sean, you were working on decentralized identity in 2019 and verifiable. Credentials, and like all these like interesting approaches, like, do you agree with that? Like, is does this need to be a global standard that is decentralized, that is kind of an on-chain identity solution for agents, or is it more it's inside individual platforms where you're setting the rules and the policies and the controls, and you don't necessarily need to have a blockchain that's enabling the identity, like, what is there a role for decentralized identity in your initial implementation of agent spending?

Sean Neville  15:30  

Yeah, I mean, absolutely, there is. You're also reminding me that, you know, my past, the internet is internet is littered with the corpses of dead identity companies, but, but I, you know, I do, I do feel, I do feel like a lot of the fundamental components that you know we're excited about with identity, and to the concept, just to recap for, for people, the concept of something like a verifiable credential, or various ways of pointing to decentralized identifiers, is just that I don't have to rope in like a Google or a de facto identity owner in order for me to prove certain things about myself, and so we've had the, we've had standards-based approaches to leveraging cryptography to do those things, so we have the tech, and we have a general set of approaches to it. Now, there's debates about have the format that those approaches can take, you know, one way or another, but generally we have the type to do it. I think that the notion of identity and AI is a catalyst to leveraging some of those things, because the other, the sort of traditional identification mechanisms we have for humans and businesses is not a really clean fit, and so cryptography and decentralized identity primitives are certainly very useful. That doesn't necessarily mean that they're all on chain, because then there's a separate piece where it makes a lot of sense for them to be on chain, but I think there's somewhat of a hybrid view in that where the hurdles to getting this more broadly adopted in the enterprise is getting a CFO to say yes, and getting a compliance department to say yes, and it to say yes, and so on, and some of the things that are necessarily in order to get them to say yes are kind of like the unsexy things that are just required to get into an enterprise, things like permission management, but in a certain way, and auditability, but with certain kinds of reporting capabilities, and things of this nature, and so satisfying some of those does mean that you need to take a little bit of a hybrid approach, and so there's multiple kinds of systems to integrate over time, though I think that the value prop of some of these identity primitives will come more obvious as we go, and some of them, even who aren't like identity geeks like we are, can sort of intellectually grasp today. If I spawn an agent and I use Olaf, and I log in with.. I'm not picking on it, sounds like I'm picking on Google, but most people do, at least, as you know, in the states they'll use Google as their login provider. So now Google knows every place that I'm logging in, which has always been, you know, one issue. Bigger issue is that as I'm spawning agents, they're all pretending to be me, they don't really have their own identifier, and so if I'm spawning multiple agents, you, as someone interacting with my agents, you just see me as the identifier. You don't see this particular agent. This agent may have too many permissions for what it really needs to do. It may have too few. If I need to sort of reject, or you need to reject access based on identity, kind of shut them all off at once. There isn't this notion of sort of fine-grained permissions attached to individual automated systems. And again, we have the tech to solve these problems. It's been complex for humans to manage, but I can manage these very, very effectively. So that's a long-winded answer, but I do agree that this is a pretty key component in making this ultimately safe and clearing the trust, the trust rule. So, yeah. So, is

Cuy Sheffield  18:38  

it fair to say to me it feels like the state of the market that we're in right now, and that you're focused on and operating, and it's almost like single-player agents that sit inside an existing org. And how do you give them the trust and the policy for me to me? You're moving your own money inside of the org, and so you want to make sure that they're doing it the right way, but it's within a bounded environment that they're just their employees that are working for the company that are doing back office treasury optimization, instead of someone having to track liquidity and kind of move it across different providers, and so the identity there, it's like it's an employee, they're never leaving the boundary, and so it's less of a like critical thing to be solved, then the next step would be, does an agent that you give some authority and ability to spend interact outside of your company, and are they going to be paying third parties or interacting in other environments where it's moving money from you to somebody else, and there it's really important. Of okay, well, what's the identity of that agent as they're perceived by whoever the counterparty is, and how is that shared if they're going to operate? Is is that the right way to think about it, that it's like internal, in-house is like the first use case, but then what is the kind of path for. As agents to leave the walls of that treasury organization and start to make payments to third parties and like what challenges and complications that that enables,

Sean Neville  20:09  

I mean, I think that's a valid way to look at it, and that's similar to the intranet versus internet sort of analog, but I think ultimately, like intranet and internet, it'll ultimately converge on the same tech stack, and we'll be using the same fundamentals, so you know there is a lot of concentration now. Is like inside my business, if I have marketing team running their work streams with AI, and I have software engineers, and I have finance, and I have all these.. what am I allowing them to do? What am I allowing them to access? What am I allowing them to do? And can I control the agents? How do I get my arms around the shadow AI that's running inside my own business and apply controls to them, and then that's a separate but related problem to once I can get a handle on my own agents, how do I allow them to interact with your agents? How do we discover each other? How do I verify it's actually Visa's agents that I'm talking to? What are we allowed to do? How can we look back and see what we did in a reportable way, which is a separate and harder problem, because we're missing some of the basic fundamentals around just basic agent tech discovery and communication. There are lots of protocols that have sort of emerged as a protocol soup to potentially solve some of these issues, but at such early days that there is no common agreement on those, and so right now I think it is very much in this sort of intranet inside my business. How can I allow my, my human capital to access my token capital effectively and safely? But the much larger, and I think related, but ultimately more valuable unlock is, how do I do that? That's in a way that spans enterprise boundaries across businesses, and you know we have solutions for this on the web, obviously. We all agreed, ultimately not initially, but we all agreed ultimately on what https is, and we even think about it anymore. We just, you can build e-commerce sites, and you just, your browser understands SSL in the background and has root certs and understands what certificate authorities are. It used to be you'd have a little lock in your browser in the old days, so you'd at least get some visual signal, and now we take it, so for granted, like, look at those things, but those things don't exist for agent to agent communication yet, and so there's just, you know, let alone for payments, but if payments are a subset of communication, then there's some just cross enterprise boundary communication, you know, sets of agreement related to identity, and also other things that need to be, I think all these things will be resolved very quickly. I'm super optimistic about this, and that doesn't necessarily need to be an academic standard, either sort of endorsed by standards, just be a de facto standard that gains momentum, and that you know that we begin to embrace, but yeah, I think your framing of it in terms of internal versus cross enterprise lines is a really good one,

Cuy Sheffield  22:42  

I think. This takeaway I was thinking about this earlier today, of it's really hard to talk about agent to agent commerce without talking about agent to agent communication, like it's just you can't just go straight to commerce. It's like right now, as far as I'm concerned, I see very few examples of agents talking to other third party agents. When I use Claude or Codex, and I put a prompt in, Claude is then spinning up other sub agents who are then kind of talking to each other, but that's all like within that single environment. The thing that I'm excited about is I have a Hermes agent. I assume you have a Hermes or Open Claw agent. How can those two talk to each other, and it feels like there's both like the underlying what is the protocol and the communication standard by which they interact, and https doesn't really seem like the right thing for it, and then there's like, well, what is the interface, and I don't know about you and Pam, but like I've been doing a bunch of experiments, like putting two agents into a telegram group, and having them like talk to each other, and it's just so hard to follow, because it's like you're a whole stream of, like, you know, there's all of this conversation happening back and forth, and then it's okay. How do you bound? Well, how many turns are they going to take back and forth? Like, you know, if it just runs away and, like, talks for hours, that's just going to burn a bunch of inference, and then if you can't read every message, how do you actually confirm what they agreed on? And so it feels like they're all these challenges to solve in just how two agents representing different entities or individuals can communicate, and however that challenge is solved, payments and commerce will be the easy part, because it's like, you know, then it's if you have trust and you have them in the same place and you have an interface to control, like you said, commerce will just be like another form of communication. Bam, how do you think about that? Have you guys done any interesting experiments around two agents talking to each other and kind of leaving the boundary of like one organization that that they're in.

Bam Azizi  24:41  

Yeah, sounds dangerous to leave to agents to talk. The infinite back rooms,

Cuy Sheffield  24:47  

that was like the crypto concept from years ago. They were just like talking, invent new things together.

Bam Azizi  24:52  

That's right. I think there would be protocols to basically set those boundaries, for sure. I think we will have the problem is now. Like, will we have a protocol? The problem is, how many protocols we will have? How we can consolidate them? Like, it reminds me of, like, the charger issues. Like, every phone had a different charger, so at some point we had to kind of consolidate all of them to one protocol, one type of like hardware. So, I think we will have some consolidation similar to early internet, as Sean mentioned, that we had like many different forms of internet, and then they all like merge, and then the protocol that everyone is using right now is HTTP. We used to have like many different protocols for communication. I think we'll have like you have one protocol that wins like 98% of the use cases, and that protocol happened to support commerce as well, so like hashtag export or two, or whatever you want to say, but that's I think where we are heading, and I think agents are going to communicate over http, so I think there would be some sort of protocol or amendment that will be added to the HCB, that's my bet, but I think some sort of consolidation needs to happen, otherwise it would be a wide west.

Cuy Sheffield  26:09  

Sean, you've seen this over decades, there's the meme of there are 14 protocols, we need one more to unify them, now there are 15 protocols. How do you think about the interplay between the application layer and the protocols in, if you are trying to solve problems, you know, for your businesses, and you're an application for agents, is that you support all of the protocols that go like a multi protocol approach, is it the successful protocol emerges from the successful application, or like what are some of the lessons in like internet history that you draw on as you think about that interplay as you build an application for agents that someday will need protocols if you're going to go to outside the boundaries.

Sean Neville  26:50  

Yeah, that's a big question. You know, it's interesting because I think it's actually more fragmented now than it was a year ago. I mean, meaning there were fewer three letter protocols to look at a year ago that were all focused on the same space. I think the thing that might be different this time, and I hope that it's not, but I think the thing that might be different relative to the earlier playbooks is companies, large as small, so incumbents as well as startups are really focusing more on protocols and application space, are not really sure where the value is going to accrue, not unified, but there was a sense that you know, in building out things like web standards for how https, I should and https could work. This was in the days of, like, you know, when America Online seemed to have more content for the average person than the entire internet did, because the internet was mostly for academics, and, but you know, sort of having a vision for what the open internet would look like for content, and for people expressing their views without gatekeepers, and so on and so forth. I think there was ultimately a bringing together at the table a bunch of key parties that normally didn't talk to each other, like Microsoft and IBM and Sun, and you know, the sort of incumbents in space who had different views on what their protocols should be, and in some cases even implemented different levels of support for the protocols that were meant to be implemented in a standard way, but ultimately came to realize that agreeing on this protocol layer, the foundational layer, will unlock a new kind of business opportunity for all of us to compete ahead, but we don't necessarily see value accruing at the protocol layer, there was a time when we thought value would accrue at the browser layer, and that's what the browser wars were about. Like, this would be the unified whoever won the browser war, it will be a winner take most, if not take all, to who controls the customer relationship with the internet. And now there's there's practically no value at all in the browser, it's it's a layer higher. And so I think that the question with this is, where does the value really accrue? Will it be the case that there's actually value at the protocol layer for owning authentic identity and commerce protocols? My belief is that those things need to be built on open standards that we can all contribute to and support, and that the value would accrue at a higher layer, but that's not necessarily agreed upon by, say, major players who are a cloud player might support 12 different things, and a Google won't agree with any of them, or whatever it may be. And so we're kind of in this early stage where the question is, how do you resolve that? I think that it will be use case and application driven, and it is the case in the early days. In order to realize those, you do end up needing to support multiple things as well as possible, which makes it hard to enforce safety and policy across all the rails, but that's the world that we're in right now.

Cuy Sheffield  29:28  

Then, within that framework, like, it feels like what's different this time is you've got the model and the intelligence layer, yeah, which is just this massive space that today, that's where the value is accruing in terms of OpenAI, Anthropic, how do you interact with the model labs and models themselves, and like, what's your thesis on the role that they will play when you're building on top of them to some extent? Of I assume that agents that you're going to enable for your clients are going to be. Multi model, like, how do you see the role of models going forward?

Sean Neville  30:04  

Yeah, I mean, I think there's even a discussion on the model, where, like, does the model sit in the front? This is an extreme example, but does it sit in the front or does it sit in the back? Does using Claude Desktop for knowledge workers replace all usage of web browsers and SaaS dashboards, and they could just build their own dashboards and co-work now in Excel, which has had a crazy resurgence as well, it's sort of tabular input alongside text and voice, and so is it like front and center, and all the things are visible in this, this very dynamic, fluid sort of LM space, or is it the other way around, that things begin to get unbundled, and we see different kinds of experiences, and LM is really more in the background, and I don't know the answer to that, and there's a lot of, a lot of discussions where it goes, but might, might be all the things in terms of our usage of models. I'll tell you where we started, and I'm not going to as far as to say it was wrong yet, but where we started was this idea that what you really want from a set of models is kind of like a kitchen full of precision tools, and so if you're going to cook a gourmet meal or bake a cake, or whatever you're going to do, you have the right tool for the right job, and you don't really want a Swiss Army knife to bake the entire meal, and so a single, like powerful large language model is like a giant Swiss Army knife, and so it makes sense to be able to choose the right tools for the job for economic and speed reasons as well as capability reasons, and so orchestration between those becomes really powerful, that diminishes the value of any one LLM provider, because you can, you have an assortment. So that was our original thesis, I think. Today the general consensus is that's not right, because you could point to Fable. Well, he used to be able to point to Fable. I don't know when this will come out, but, but you know, Fable was on the table briefly. We're Opus Four Eight or GPT Five Five, and they will do everything, and it's really just a matter of processing power. And so, if you can afford it, and you can throw the processing power at it, then that Swiss Army Knight is your kitchen full of precision tools, so I'm not willing to say that we were outright wrong, but that's the way things look today, is that these models are so incredibly powerful that you don't necessarily need that level of, it was a time when we were fine tuning the early, you know, the early models, and that proved to be mostly a waste of time and money because of the advancements in these other models, but I would imagine more a world where we're doing specific things like setting up a harness with all the data and all the contacts and all of the proper connectivity to tools in the light for say investment management, well, that's a very different structure compared to things like executing effects markets, and so those things require certainly different kinds of ways of looking at data, and aren't folding them into the right context to be effective and reliable, but it is also even though it looks like today that the intelligence is not the bottleneck, and that you could just rely on the intelligence that's available, that it will be the case that you'll want to, you'll be advantages in choosing a tuned model at the intelligence layer, as well.

Cuy Sheffield  33:02  

Yeah, it's super, super interesting. And so it's almost like the harness in that scenario. What I hear is like the harness is actually creating the most value, and the difference between

Sean Neville  33:12  

it is today, I think, absolutely, the

Cuy Sheffield  33:15  

difference between using just one model to try and run your treasury management strategy versus using a harness that has the right model with the right tools with the right data sources, and then it feels like, isn't there also a big cost consideration? Bam, of I feel like a month or two months ago people were in this honeymoon phase of, oh, like these models are super powerful, I could build anything, and then now all of a sudden it's model orchestration, token ROI, like people burning through their budgets. How do you see that playing out? And like, if you're building use cases for agentic commerce or agentic treasury, do you need the most super powerful frontier intelligence to tell you when to move $100,000 from this pool to this pool when the rates go up, like, is can can you do that at like a better cost per token? Bam, how do you think about that?

Bam Azizi  34:07  

They call it token hangover, right? So during so much of token that now you're hangover. So as far as agentic commerce, I'm less concerned because the level of like intelligent you need to transact when it comes to the final, like, or the last mile delivery, it's not that much, so whether the agent is super intelligent or just a normal agent, you want the rail to be there to enable them to be able to pay or do things or transact the way they want, but I think it would be important from that perspective, that this agent be able to communicate with one another in the most efficient way, so the way that you've mentioned that, like, if you just leave two agents without any boundaries, so they start talking and spending a ton of tokens for just like spending like two cents, you really want to kind of have the. Boundaries to make sure that there is no diminishing marginal return on those transactions, so that's that's from financial perspective. From like intelligence perspective, I think I agree with Sean that their thesis still holds water that we'll have different agents, so for instance I'm seeing a lot of agent to agent transaction is because one agent, for instance, wants to buy a pair of shoes, so it's basically built the way, or the most efficient way to transact, but also is not the best agent to read all the reviews, like what type of shoes should I buy, right? So they would go and pay another agent, and they spend like $10 or maybe a fraction of $1 to read all the reviews and come up with an idea, like what is the best shoes for this specific agent, and if that similar general like purpose model like agent wants to do the same thing, it has to burn more than like $10 token or $100 token, so it makes sense for that agent to just go and buy that versus just basically do it itself, even though it has the capability, but from our perspective it doesn't make sense to burn all the tokens to just buy like a $10 or $20 shoes, right? So that's that's how I think it will basically force us toward not having one general purpose model agent will have like multiple agents that they're good at what they're doing, and it's basically traditional like buy versus build, right? So buy versus do since stuff us, me doing it, which cost me under dollar, would go and then pay the $2 to this specific agent, and that agent will do that for me, and that, that will basically increases the number of transactions we'll have on the agentic commerce, so it forces us toward that, that word, I think. First time, Jeremy, co-founder of Circle, the other co-founder told that they like the agentic commerce is important because humans, they do two transactions a day. Agents can do 2000 transactions a day, or a second, and we have 5 billion human beings connected to internet. We will have hundreds of billions of agents connected to internet. So that's that's where the agentic commerce boom will happen. From my perspective,

Cuy Sheffield  37:17  

I want to come back to the agentic treasury use cases of just kind of where the market is today. Who are the types of companies that you see as like the ideal customer for this that should be experimenting and using it right now? Like, isn't a prerequisite that they start with they already have stable coins on their balance sheet, and so it's a function of you have stable coins, and then now you're trying to optimize the yield on those stable coins. Is it a company that has a fiat treasury that wants to optimize the fiat treasury, and you're doing it in fiat, or you're taking a fiat treasury, converting the stable coins, and optimizing? Like, what? What is the type of business that should be working with you all to implement and start to use a Gentic Treasury that you think it's going to be the first wave of success cases that then make this a real category that more CFOs and treasurers start to gravitate towards,

Sean Neville  38:11  

so ultimately I would say we think our end users will be agents themselves, we're just not there for all the reasons we've been talking about, we're not there yet, my agent can't even find your agent and safely talk to it, let alone let you KYA and onboard to a bank, but we do think that's where we're going. I mean, that's why we're pursuing the regulation that we're, you know, we're pursuing at the same time we're building up the product. But today, for those who are beginning to build these agents and begin to build these automated systems, it's either the case that we sort of found they're either there's trying to two axes, stablecoin native and AI native, and the mixture of both is perfect for us, but usually companies are more one than the other, and so on the agentic side, there are people that are beginning to connect automated workflows for various use cases, but you just say maybe it's for things like paying suppliers, for instance, in kind of a general use case, and they have a number of underlying rails and accounts and existing mechanisms for paying their suppliers in different currencies, and as they begin to automate it, they run into problems around just integration. How they connect it? How does an agent manage all these things? How does an agent turn dollars into euros, and you know, euros potentially into backing to stable coins, and then you know, and so on, and so forth, and so a lot of it is being able to do that, and then being able to do it safely, so that the policies that they, that they enforce, and policies, meaning things like counterparty rules, or trading window rules, or or timing and velocity, and you know, these sorts of things could be sweep rules as well, in the treasury case, so how do they do that effectively in those sorts of cases, we are able to answer them agentically, so agentic workflows are able to handle all the integrations, including clearing and sweeping between different stables, and on and off ramps into fiat, and in understanding how a particular payment needs to be made, including currency conversion, but really the layer that we're focused on most is how do you do that safely in a non. The way across all the rails, stable coin included, on the stablecoin native side, it's more that companies have kind of leaned into stables, usually one of two that have significant liquidity today, and so they find themselves a kind of an idle flow, an idle cash flow problem, which is that if you have dollars typically sitting on in a corporate treasury, then you get at least invested in short-term treasuries, but if you have USDC, for instance, it's unclear if you're traditional in the traditional treasury function what you do with that. You're not trying to get VGN returns, you're not looking for, like, how do I optimize across all my vaults? It's more like, how do I get just the equivalent of my 90 day treasuries that I used to have, so they get my three and a half percent, or whatever it is today, and so in that case, because we are able, our sort of agented framework is able to handle multiple rails, it's a combination of crypto, you know, lending vaults, but also treasuries and money market funds and traditional assets as a blend of the assets, so yeah, I don't know that exactly answered the question, but usually it's companies are not really AI native, and also stable coin native. If they're AI native, they often don't even want a stable coin, they just want $1 to work effectively. And so stable coins are the solution for that in many cases, but they just want to view that as $1 And so then the clearing capability becomes really important.

Sean Neville  41:15  

So if tether and USDC are trading against each other, and 198's cents and one's $1 one, they don't want like mark to market mechanics, they just want to look at it as $1 and so we can solve the sort of singleness of money problem for them on the back end, and on the other hand, if people are looking more like I'm trying to automate the system, but I really needed to be auditable in this way and hand a report to my board, so it understands the intents that were executed by this, how do I do that, we can solve that problem too with the governance, the governance plane.

Cuy Sheffield  41:43  

How is it different between, like, fiat only versus let's say stable coin only, to just like make it very simplistic? Because in a stable coin context, you can give the agent a wallet, and then you can create a set of policies around what that wallet can do and enforce those policies, which feels like that's one of the coolest things that stablecoins have. What's the equivalent of that with traditional fiat accounts? And how do you manage? Like, is it just for you all it would be much more convenient and more flexible, and there's a lot more that you can do if it's on chain, which seems like that's Pam's thesis, or is there a world where even if a company says, "I refuse to have any exposure to any stable coin, I just want to do everything in fiat, but have the policies and the auditability for my agents to move money? Is that like, okay, we could solve that too, but it's harder. Where does the kind of unique aspect of stablecoins come in at the policy layer.

Sean Neville  42:43  

Yeah, it's concentration of trust answer, really. So, as you think about four layers where policies are enforced, one is just AI guardrails, which are almost like suggestions, frankly, because there's - they can be escaped so easily. But I think that space will mature as well. So, there's an AI guardrail layer, there's a policy engine layer where if you put all the capabilities in the policy engine, you're trusting whoever runs that engine. In reality, there may be multiple engines, but then still it's a federation of vendors who you trust. So, so there's that layer. Then there's the cryptographic layer in the wall. It's like if you can encode policies in a TE, which is how we manage our underlying stablecoin infrastructure, then our policy engine can't possibly move those funds, but neither can your agent. They have to agree. There has to be based on how you define your policies, we can sort of enforce this cryptographically at the wallet level, so you don't actually have to trust your policy engine because of the way cryptography works. And the fourth layer is on chain itself, so there may be rules that are encoded in a certain kind of transaction that your agent is attempting to execute, even if the policy check is done, your agent says yes, the guardrail says yes, it may still be that with this kind of what you're trying to do on chain, so this is the kind of programmable on chain representation, it still is not permitted, so there's if you kind of look at those four layers, if you take off the on chain and the cryptography lane, you're looking more at like custody, you're just concentrating trust in two layers, the guardrail and the policy, whoever's running the policy engines, and you know various enterprises may have different views on that. Typically, the way to things like, how do you, how do you trust a policy engine is sort of like, well, they're regulated the same way, and so there's a regulatory framework to say if they violate their fiduciary duty, then there are implications for that, that are enforced by the sovereign who has ultimately the ability to carry guns,

Cuy Sheffield  44:30  

and so it's ultimately to me it's a trust problem. I think you know where I would see it is why I was so excited about building Circle and creating stable coins. I would like to see trust encoded in technology rather than be left to humans and the businesses that the humans create, but certainly not all businesses have the same view. Some have the complete opposite view and would rather trust the sovereign and legal contracts written in English enforced by judges, and so it's both, and trust gets distributed across them all, so. So, if I'm hearing you correctly, it's there's an argument that the more trustworthy way of having four layers of trust for a genetic treasury movement is doing it on chain with stable coins, because you still have, you can have off-chain kind of legal enforcement and guardrails, but if you're doing it purely off-chain in traditional fiat, you miss some of the cryptographic verification that you can have when you have a wallet, when you have a policy engine that runs in a TE, and like that just doesn't really exist in traditional financial products, which is fascinating. That's like, if, if the goal is a genetic treasury, even without realizing it or thinking about it, stable coins are better positioned for agentic treasury to work in a more trusted way than traditional fiat, and, bam, that's it. Seems like that's consistently what you were saying, is like you think that it's the smart contract and the functionality you can build into on-chain assets, like, is is that kind of your, your thesis as well? But agentic drives the need for those properties that stable quits have,

Bam Azizi  45:59  

that's correct. I think every asset will come and chain, and every logic will come and chain, and this specific logic and rule-based decisions are no exceptions. So, I think that, like, blockchain smart contracts is going to run that layer.

Sean Neville  46:14  

I agree. I think they're catalysts for each other. Honestly, I think there's certain kinds of agentic commerce that are not possible without cryptographic and on-chain solutions, and then vice versa, as AI commerce really begins to take hold, there are catalysts for each other. As it increases the adoption of blockchain solutions, more blockchain solutions we have, the more AI solutions we have that can leverage those effectively. More, there's it, there's risk still involved in on-chain solutions, of course. The risk is just in a different place, and it comes down to things like where you manage the signing, and how do you, how do you handle your key management, and things that are also can be addressed, but are not typically the way an existing enterprise finance department does things today, and not because it's not a good way to do it, but because it doesn't clear the hurdles that you need to clear in order for it to be approved in the enterprise, and so I think the more we see the power and the value of AI, the more that these other infrastructure, sort of blockchain infrastructure hurdles, will be cleared, because it's not possible to do the first thing without the ladder, and so they work, they work together.

Cuy Sheffield  47:11  

Yeah, it's a great way to end. We could talk for hours on this topic, and we're gonna have to have you back in a few months. Is just the pace that you know the entire industry changes. So, thank you so much for listening, Sean. Where can people learn more about you? And container,

Sean Neville  47:25  

container.com is a good place to keep tabs. We're on all the things that has all the links,

Cuy Sheffield  47:32  

and Bam, what about you? And Mesh,

Bam Azizi  47:34  

Mesh Pay on LinkedIn and x.com or Bam, as easy Mesh on X, or Telegram.

Cuy Sheffield  47:40  

And you can find me on X at Kai Sheffield and visa.com/crypto So, if you haven't already, please subscribe to Tokenize on Apple, Spotify, or wherever you get your podcasts. If you enjoyed this and you want more, leave us a review - it really helps others find the show. So, thank you, Sean Bam. Always a great time talking to you both, and we'll have more soon.