Value Accrual in the Agent Economy: Visa's Head of Crypto on Stablecoins, x402/MPP, and Card Networks

Transcript

Gabe Tramble (00:01) So Kai, tell us if stable coins are internet native money and agents are basically the internet native buyers, then where does the value accrue in this entire equation and and stack?

Cuy Sheffield (00:16) Yeah, it's it's the trillion dollar question. I think it's so early that you know nobody knows. But I I think a few principles that we have here are we believe stable coins are amazing next generation infrastructure. And you know, the ability you know to have blockchain networks that can move value, you know, twenty-four seven across the world that are programmable, you know, that have a low barrier, you know, to entry to be able to spin up a wallet, interact with. Yeah, I think is fundamentally amazing for you know the increasing the velocity of money, bringing more people into digital payments and the financial system. I think all of that is is incredible. But on the other hand, we don't look at stable coins and blockchains as payment networks. Payment networks are fundamentally different than payment infrastructure because a payment network, it's it's not just about the technology. It's not just about the protocols, it's about the governance and the risk and the rules and the controls and what happens when things go wrong. And I think that's where, you know, when we look at at protocols like X402 and MPP and we're incredibly excited about them and we think that they're going to play a big role. We don't think that the the protocol themselves is complete as a payment network. We think that there need to be layers built on top of these protocols, you know, that can actually commercialize them and bring them, you know, to many, you know, more people and and customers. And and I think today, you know, by all means like a gentic commerce just doesn't exist like at any scale. Like we've got like the the bones and the kind of the early foundations, but there's just so much more to be built on top of it. And so, you know, I think that there's a a massive opportunity. you know, for new payment products, ⁓ you know, to be able to to take what X42 and MPP have as a protocol sense and be able to commercialize and enable kind of networks with governance and trust and structure, you know, that can actually move value on top of those protocols. So ⁓ that that's my my initial take.

Gabe Tramble (02:28) In okay, so this this network layer, right, which which you're saying is is native to almost like these fiat card systems, right? Wha can you kind of differentiate where the value is in cards and then where the value is in stables because there's like the the the way that people are thinking about this is it's almost one or the other, but you guys are at an interesting intersection where you're adopting both at the same time. And yeah, can you just help us understand more this this network versus ⁓ infrastructure piece?

Cuy Sheffield (03:02) Yeah, I I think anytime people kind of position it as you know stable coins versus cards, it's just such like a a zero-sum thinking at the very beginning of this like you know potentially enormous new space. And and we very much believe it's it's and and that you know both stable coins and cards will play a major role in the future of agentic payments. I think there are yeah perhaps you know different use cases and different client segments, but like increasingly we believe stable coins are one of the best backend infrastructure pieces for payments. And cards and networks like Visa are gonna continue to be one of the best front end ⁓ payment networks you know that both consumers and increasingly agents you know in the future will interact with. ⁓ and so I think like the back end piece of of what stable coins do is they actually enable the money to move and you know frictionless you know movement of money and being able to do it, you know, 24 seven, not having bank holidays and Doing it in a programmable way is is great. But I think the the missing piece is as I mentioned before, it's it's the trust in the governance. Yeah, it's knowing that, you know, if I'm gonna transact with a merchant, someone has vetted that merchant, someone is taking responsibility for that merchant. If something goes wrong, you know, there's going to be, you know, either some way that I can, you know, get my money back. Or some way that prevents that merchant from continuing to scam other people. And so I think there's both this trust element as a consumer. When I see a merchant that accepts Visa, you know, I know one, I've got an expectation the transaction's gonna work. Like that, that's the first thing. And then two, I know that that merchant has been vetted. And I know if something goes wrong, like there's going to be a set of clear rules and processes in place, you know, to to make that right. And I think what that does is that lowers the mental transaction cost. Where if I see something I like, I just buy it. I don't have to think about the payment. It fades into the background. And I think that the challenge that stable coins have right now, which is just kind of the general property of crypto, in being an open public, permissionless network that anyone can interact with, it doesn't solve that trust problem where you know, take X402, for example. Anyone can create an endpoint on X402. And so if your agent is looking to transact with a merchant, You know, they can find an endpoint. Well, what if that endpoint is run by a company that you don't want to interact with, that you don't want to pay? And you know, maybe that's yeah, someone that's in a sanctioned you know country, maybe that's somebody that you know is unlawfully you know selling what those goods are? Like there's no nothing stopping someone from standing up a merchant endpoint. What if I buy something from that endpoint and I don't get what I bought? What could I do? What prevents that you from happening again to somebody else? ⁓ what if that merchant endpoint?

Gabe Tramble (05:54) Mm-hmm.

Cuy Sheffield (05:57) Isn't even selling a good or service? What if that merchant endpoint is just malicious software that's trying to get me to interact with it? So it can prompt eject me. Like there are all of these things that when you have an completely open permissionless ecosystem, while it's great that it brings people on and enables like, you know, any individual could stand up a merchant endpoint. I think it's really important if you're actually gonna scale commerce to know that you can trust that merchant endpoint. And so that's what we're really focused on of like, how do you add a trust perimeter?

Gabe Tramble (06:21) Uh-huh.

Cuy Sheffield (06:26) And kind of a trusted layer and network that can then sit on top of those protocols. And I think that that's the role that that Visa plays. And then, you know, on the other hand, you know, I'm a crypto guy. I've been in crypto for for close to 10 years now. And like I get nervous about giving agents a stable coin wallet. There are just fewer controls that are in place. And an agent managing a private key and like you know, what happens if if things go wrong, the money's gone, it's a bare asset, like you know, there's nothing I can do. Being able to give an agent a card.

Gabe Tramble (06:37) Yeah.

Cuy Sheffield (06:55) We believe is a better experience because you can have controls at the CLI wallet layer, you can have controls at the issuing layer, you can have controls at the network layer. And so you just have these multiple lines of defense, you know, that prevent this scenario of you just lose all your money, which I think we're in this point where it's the frontier, people are trying a bunch of things out. But if this is actually going to go mainstream, there's more of a need than ever for trust and for a place that you know the merchants can be trusted and that you know that.

Gabe Tramble (07:20) Mm-hmm.

Cuy Sheffield (07:25) If something goes wrong, you know, there are protections in in place to pre

Gabe Tramble (07:29) Yeah, yeah, that that's that's interesting. And you and you guys have basically decades of this trust to to lean on to and ⁓ at the same time like embracing these new novel primitives like X World Two and and some of the other ⁓ agenc commerce infrastructure. The with that in mind, if you were not at Visa and let's say you're a builder, knowing what Visa is doing in the space, ⁓ what would you not build?

Cuy Sheffield (08:00) What would you not build? Interesting. I think a lot more about like what what people you know should build. ⁓ I I think it's it's very, very hard to build new payment networks and to build that trust and acceptance that it's taken us yeah decades to do. And so yeah, anytime someone says, Okay, I've I've got a better technology, it's it's faster and cheaper and it's programmable, so so we're gonna you know displace Visa and like Visa's value, you know. Only a small fraction of it is the technology. Like it's how do you coordinate and get you know the largest entities in the world to align on a single set of rules and standards and you know to cooperate around those while still competing, you know, with each other. That's a a ⁓ much, much harder thing to do. ⁓ and so I think that that the trust and coordination piece is not something that, you know, a startup that is you know a single entity. you know, is able to to really solve, you know, at at a global scale. I I think it's it's makes a lot more sense to be on one side or or the other of the network. I think there's a massive opportunity in building new merchant endpoints ⁓ and creating what are the next generation of headless merchants. I think you could take a look at pretty much every you know SaaS service out there, and you know over some time, many of those services are are gonna have to go headless. They're gonna have to become accessible And be kind of thin, super lightweight, easy endpoints that you can then purchase inside of any agent harness that you want to be in. And I think the interesting question is gonna be, well, are some of those SaaS providers gonna kind of decapitate themselves and and go headless and say, All right, you know, we're gonna meet customers there and we're seeing Salesforce and and some others. But I think that there's SaaS in in the kind of the software market is so big that there will be plenty of incumbents that you know they just can't operate in that structure. It's just it It's not the core model that they were built for of, you know, seat based pricing and enterprise sales and like fancy dinners and like that. That's like a lot of how software is sold. And so I think there's this huge opportunity to create, you know, valuable software products, ⁓ that are just built for agents, you know, on day one that are these headless merchants that are accessible, you know, over those protocols. So I think that that's like one of the biggest opportunities, you know, right now. ⁓ and and then I think the other piece of it, which curious what you've seen is a lot of the X402 endpoints today, I think are like a a single tool. And I think at the end of the day, a lot of it's a lot harder to sell a tool than it is to sell a workflow. And so I think it's really how do you combine together and compose different services and different endpoints to create something that is a lot easier for people to understand and consume. And I think we're still in this phase of there are thousands of tools

Gabe Tramble (10:47) Mm-hmm.

Cuy Sheffield (10:52) But a lot of people don't really know how to use them. And so can you actually, you know, even if it's starting with the existing X42 endpoints out there, create a product that combines those endpoints into an actual workflow and then being able to sell that workflow? I think those are two of the biggest opportunities right now for earlier stage companies to be able to come in and have an impact.

Gabe Tramble (11:13) That's super interesting. Yeah, I I I definitely agree with you. The the X the at least the value that I've seen for X402 right now, today, is mostly on the prospecting side. So if you're trying to, you know, find emails or customers or phone numbers, et cetera, the value is there, but it's still super long tail. And and to your point of the the opportunity, I know ⁓ Merit Systems who who makes agent cash and ⁓ Locus Pay. Locus Pay like a YC pa agent pay provider, same as ⁓ agent cash. They both are basically bundling ⁓ endpoints to have like these specialized ⁓ ⁓ services. So if it's like, you know, social media services, then it bundles like all these different social media platforms or you know, prospecting and it bundles, but I haven't seen the workflow itself of basically ⁓ giving the agent a skill, I guess. So they are doing the skills, but there's no kind of like robust packaging of this that that I've seen so far. So that's pretty interesting.

Cuy Sheffield (12:16) Yeah. So I think that's super consistent with you know I've had a bunch of conversations in the last two weeks ⁓ with you know different merchant endpoints, merchant endpoint providers. And I think that that's a a common theme of you know, it's like go to market in sales, you know, prospecting, generating leads, like that. That's one of the areas that there is some kind of real, very early organic traction. And so I think it's it's a natural direction to go to try and stitch together endpoints into a workflow or something that a salesperson you know can use. So I think that's on the supply side. On the demand side, I think right now that's that's kind of the bigger problem, you know, of just there's not a ton of easy demand you know coming into these these endpoints, despite there are some like very useful endpoints. And so the question is like, how do you a hundred X the growth of the demand side of people that can access and and kind of pay as they go? And I think there are a few different approaches that are are being tried. And I think one approach is we'll just create new front end interfaces. And you know, I love what Meredith is doing with Poncho. I think that's a super interesting interface that is like a harness that's kind of built in of being able to to run these workflows. Tripstack and others, they have their own front end. And so they're there are a bunch of people trying to like create their own front end to then drive demand into their endpoints. ⁓ but I I think that one of the challenges is like you have to abstract crypto away, you know, to be able to to get to you know, a number of mainstream, you know, developers and users. And so we're spending a lot of time just trying to figure out, you know, how do you both enable someone to just put a card on file and be able to access, you know, any of these, you know, endpoints, at least the ones that, you know, we think are trusted and vetted and and can be, you know, good services. And then how do you get that embedded inside, you know, every agent harness out there? and I think we're still in such an early stage where most of the agent harnesses, like you can't Pay for anything outside of the harness. It's like the only thing that you can do inside Cloud Code is pay for anthropic credits. You know, the only thing you do inside Codex is like, you know, pay for open AI credits. And I think it's a matter of time before you see both a number of emerging competitive harnesses, and whether that's you know open source ones and Hermes and Open Code and others that are gonna make it easier to access like the broader universe of third-party models and tools, you know, natively inside that that harness. Yeah, as well as eventually the larger ones kind of opening up to that. And so, you know, I I think that the supply is further along than the demand. ⁓ but our goal is just how how do we grow the overall ecosystem and and how can Visa use our platform, our distribution, our trust and and just the fact everyone has a Visa card, you know, to be able to get more people having a card on file in a place that could then access this kind of growing universe of of endpoints being created.

Gabe Tramble (15:04) Mm-hmm. Yeah, or from the from the enterprise side, just your work so far and and you talking to these these merchants. And and just to be clear, there's the The user, the end user is probably an individual or maybe a company, an individual at a company using an agent like Claud Code, right? That has the ability to pay for services for the service providers themselves and the folks that you're talking about on the enterprise side. Are you what are you noticing regarding their kind of behaviors and thoughts about ⁓ agents in commerce? Are they kind of adversarial? Are they nervous about, you know, bringing their products onto the platform for agents? what what are you seeing so far?

Cuy Sheffield (15:44) I think there's a big separation between you know developer facing merchants and enterprises and traditional kind of consumer facing merchants. ⁓ I think on the consumer facing merchant side, you know, it's just it's a lot earlier and there's just a lot more questions. And like there there's there's like a FOMO of we like we we want to be involved if people start using their agents to help them shop. But there's there's like real risk of you know, a lot of the value of a lot of merchants is their storefront, you know, their website, you know, their their experience, their app. Like, and and I think that there's there's this question of like, okay, well, what does it mean to open up and and let an agent be able to transact with me? Yeah, how do I control the end experience and the brand and you know, how do I decide which agents? It's much easier to say, let me just, you know, try and embed inside Chat GPT and not have like anyone's hermes like showing up to my. my store. And so I I think that's one of the reasons we've we've been focused on kind of this this idea of like the the trusted agent protocol of like, you know, how can you have some level of identity, you know, to an agent and know who the agent is and be able to have a signature that, you know, they've they've been vetted you know by Visa when they show up to a merchant. Cause I think right now it's it's either kind of closed loop, you know, get embedded inside chat GPT, which isn't easy to do. And like, you know, that there's been kind of stops and and starts to that. Or it's you're letting everybody in. If you're letting everybody in, like, you know, what does that mean? You don't, you know, have a lot of control. And so I I think that there's there's a lot of work to do to find kind of more of a middle ground, you know, I think protocols like UCP are are playing a big role and like that's still early and and I think you know has to to to get built out more. ⁓ on the developer facing merchant side, I think there's just a lot more willingness to experiment and to say that like every developer is using a coding agent. And I need to like meet them where they are. And yeah, I need to be able to have an interface that they can interact with me. I think one of the challenges right now is that the way X402 was built is it's kind of like a it's a guest checkout. You know, that that's the only experience. It's like every agent that shows up to an endpoint, you know, gets the same pricing, the same product. Yeah, endpoints are very static. ⁓ and it's hard to like you know really understand who that customer is. And serve that customer really well. so I think we need to see more of an evolution towards dynamic endpoints. You know, these merchants, like they they want to know who's the customer, what's their what's their email, be able to communicate with them. You know, in the future, maybe there's you know different pricing or different endpoints that would be exposed to some customers and not others. and so I think that there's so much more that could exist in terms of enabling not just a guest checkout kind of one-time experience. But a like true headless, you know, any agent inside any harness being able to pay for any developer service, ⁓ but having the developer service still have a relationship and be able to interact with that agent. I don't think that really exists yet. ⁓ but I I I think that there's it's more likely of like I like to ask questions of like what where is the first a billion dollars of you know agentic commerce like going to be spent? ⁓ and I think that, you know, one, I'd argue like,

Gabe Tramble (18:55) Mm-hmm.

Cuy Sheffield (18:59) There's already billions of dollars that's spent of agents buying tokens. It's just an agent, you know, inside of a single platform buying tokens of a single platform. And so when are you going to have like a an open billion dollars spent of an agent inside one platform buying something from another platform? And I think it's much likely to be inference and kind of very developer-facing, kind of agent-facing services than it is to be shopping assistants like in the near term. So that's why like I'm very focused on the developer space.

Gabe Tramble (19:27) Got it. Yeah, okay. And and and the reason you're focused on the developer side is because there's already one demand. There there's some ⁓ visual of demand there and they're more receptive on the enterprise side.

Cuy Sheffield (19:41) Yeah, I I think it's it's the best use case for an agent to date. Like, you know, what are people using agents to do? They're using agents to code and build. Like, you know, that that is by far to me the the area of you know just agents that are doing productive economic activity. Like and and the scale at which agents are coding is so much greater than the scale at which agents are serving as shopping assistants. And so I think that whatever protocols and networks and products and solutions that can help solve for making it easier for agents to pay for the things that they need as they code. I think that that's like the the beachhead and in the first place that agentic payments really happens. And then once an agent has a really good way to pay for an open marketplace of you know services it needs as they code, it probably makes sense for that also to be used to pay for you know ordering you know DoorDash or whatever it is that this developer needs. In addition. and so I think it's it's a lot of the other Egyptian commerce could end up being built on top of you know what the developer facing use cases are. And ⁓ yeah, we'll see. But I we just want to be in the ecosystem where people are using agents and and that's that's developers right now.

Gabe Tramble (20:53) Yeah, right now I th I think there's like this looming overhang. And maybe people want to like pull the curtain back and and look at it at at the ghost or the boogeyman in the closet, but I kind of get this feeling that there are these large incumbents, you know, sim similar to yourself ⁓ at Visa or like Stripe or even Anthropic themselves, right? or Cloudflare. There there's like these huge infrastructure providers or network providers. That can just like turn the lights on. And in I would say like for a builder, what's scary is if Anthropic just, you know, works with Cloudflare and then they just enable wallets to purchase on like all these different websites, right? ⁓ do you think that path can happen where just, you know, ⁓ in in a month we just see like all these transactions because two large institutions just turned on the lights? Or do you think it's gonna be more of this kind of movement or like slush of of supply and demand and and activity to kind of ⁓ a l a longer term thing of of agentic commerce.

Cuy Sheffield (21:58) Yeah, I I think personally it's it's much more likely, you know, to be this kind of back and forth and like, you know, bottoms up, you know, some emerging new use cases from new players. And then that, you know, people learn from and then the large players kind of adapt as it's I think it's very rarely in new markets, you know, like this, yeah, something that, you know, it doesn't exist and then overnight in a month and now it exists like at scale. ⁓ and I think everyone is is kind of

Gabe Tramble (22:21) Yeah.

Cuy Sheffield (22:26) learning from each other and experimenting and and testing things out. And and I think in in many ways it's it's easier for small players, you know, to experiment and be nimble and kind of find, you know, the early signs of of where there's demand. And then you know larger players you know come into the space after that. So yeah, I I think it'll be it'll be a a give and a take. ⁓ I I think the first market to me that is going to be really, really interesting to watch is just like what happens with you know token share on inference.

Gabe Tramble (22:44) Mm-hmm.

Cuy Sheffield (22:55) Which I think is gonna be one of the mo the the the most important trends like over the next you know six to to twelve months. And like the the feeling that I have, and and and I don't know if if if you share the same thing, is it's almost like there was this this phase of like coding agents emerged, and you know, Claude Code, you know, to me was like that was my moment of like, whoa, like I can build anything with this. Like this is incredible. It's like the best product I've ever used. And and then there was this like honeymoon of like everyone just vibe coding, just like building anything and

Gabe Tramble (23:18) Yeah, yeah, yeah.

Cuy Sheffield (23:24) I'm spending hundreds of dollars a day and I don't even care because like I'm like coding now. And it's so you go from this like token maxing mentality that I think was many very much driven by Claude Code and Codex. Like those are the platforms that people in are in there using and you know are experiencing the magic of having a coding agent, towards then waking up with a bit of a hangover saying, How much did I spend on tokens over the last few months? And like, what do I have to show for it? And like, was there ROI on that? And like one thing I I was just talking to to some people about today is if you think about inference as like a new merchant and like product category. Like it's a product, it's a skew, you're you're buying inference, you're buying intelligence. There aren't that many analogs of you don't know the cost of it until after you use it. And so it's like you show up to a store and you're like, okay, I I'd like to or service, I'd like to do this. And they're like, okay, like, you know, we're gonna do it, and then we'll send you the bill for what it costs after. And you're like, well, like I was just trying to set up an agentic workflow to like download some podcasts. I didn't realize I was gonna be a hundred dollars that like this agent got like stuck in a loop and it was doing all these things it probably shouldn't have done. ⁓ and so I think just the the the kind of push towards how do you justify the ROI and be able to measure and track, like me personally, like I want like a mitt.com for token spin. ⁓ it's just like I want to know everything I can, you know, about How do I lower the cost of my token spend, getting as much value and like measure that ROI? But then you have this like structural conflict that, you know, if you're anthropic or if you're open AI, like it's not in your interest to like make it super easy for people to like reduce and optimize their token spend by routing away to like cheaper models. You want everything to be going through your models. And so then you see like, you know, the open source harnesses, you know, start to emerge. And I think companies like Open Router are building amazing products to have. Like a single integration that can then give you access to 200 plus models. ⁓ and then does that get embedded more into these different harnesses? And yeah, I think you could do that on Hermes now, you could do that on open code. ⁓ and so I think you'll you'll see more of this decoupling of you know agent harnesses that are like model agnostic, that are like built to be more ROI-driven, ⁓ and then access many models and tools. Instead of this first phase of the market where it's like the harness and the model, they're kind of coupled one-to-one. And anytime you're using that harness, you're using that same model. And it's just very hard to justify like buying frontier tokens for very basic coding tasks. But the problem is that it hasn't been easy to like orchestrate and like have access to a broader universe of many, many models that are out there. So I think that's one of the first use cases where like again, injective commerce is happening of agents buying tokens. They've just

Gabe Tramble (26:13) Mm-hmm.

Cuy Sheffield (26:14) Been in this kind of environment where they're only buying tokens from one store that has one price. If you give agents the ability to help optimize and have services and products that let them buy tokens from many stores, they could then optimize the prices that they're getting and the value from that. And so I think that's one of the first areas that I could see in. And I think that the the total value and the amount of money that's spent on agents buying tokens is likely to be bigger than agents buying everything else combined. 'Cause it's like that's the lifeblood of like an agent has to buy a token to run. And so sure, it's gotta buy some data, it's gotta buy some some other things. But like if you can really figure out how to make it easier for agents to buy tokens, I I think that's gonna be a a really interesting space.

Gabe Tramble (26:48) Yeah. Yeah, wow. I I I really like that illustration. I'm gonna try to summarize. So so basically a a huge opportunity you see is agents buying their own inference and or or mu or ⁓ kind of like energy into stimulating the harness and the model. And right now you're saying that there's like this opa opacity where it comes to ⁓ pricing and and only one ⁓ customer or one one merchant which is like anthropic or or ⁓ open ai and with these open source models they're giving you all these different options And the new harnesses like Hermes, I know they allow you to basically adjust the model depending on the task. So essentially you're saying if the agents have this open marketplace in which they can be more efficient in selecting the correct models to use for the correct tasks, then you think that right there is one of the biggest use cases for agentic commerce.

Cuy Sheffield (27:55) Yeah. And I I think this concept of model orchestration is really interesting. Of like how how do you find the right task for the right model? And I think there are only going to be more models that exist that could be specialized and potentially better at certain tasks, but could also do those tasks, you know, at a lower cost. But the problem is today it's pretty hard. Most people who are bive coding don't have, you know, a router that's automatically set up in whatever harness they're using that is automatically picking the right model for the right task. I think that's kind of like an inevitable direction that it should go. And that I think developers, you know, should want, you know, some type of intelligent router that's you know helping to pick the right model for the right task and having access to an inventory of many, many models that are that are out there. And I think that that's going to be one part of how do you optimize token spend and and justify the the cost of it, the ROI. ⁓ and to me that that's a that is an agentec commerce use case. It's it's an agent that is buying inference. Or also buying data and tools. And you know, I've I've just been like, as I go down the rabbit hole, like there's so many things I'm learning where, you know, there are plenty of times where if you have access to the right tools, you know, you don't even have to you know pay for inference, yeah, because yeah, the tools can go and kind of do certain types of work and store that data, you know, locally that you could come back to instead of having to run it, you know, through a model. And so I think there's so much room to optimize just the core like. back end guts of how an agent operates in writing code. And I think that if you just unpack like, you know, agents are going to write more and more software in the future, if that goes from, well, the agent spends at one merchant of the platform that they're on, to like the agent is spending at thousands of merchants and being able to access these different tools, like, I think that could be like one of the early agentic commerce markets that that's very interesting to look at.

Gabe Tramble (29:51) Interesting. Like self fill up or like self charging for ⁓ for vehicles.

Cuy Sheffield (29:55) Yeah. And I I think more competition of being able to because if if you're building a new model, you know, you might have a model that's really good at a certain task. How do you then go and reach customers? And so it's it's like this new two-sided network problem where it's like you have agent harnesses on one side and then you have you know models and tools you know on the other side. And in an ideal kind of efficient universe, any agent inside any harness should be able to, you know, access, discover, and find. you know, any model or any tool that could help them do their job and be able to pay for it as they go. To me, that's like that's how it ought to be. Like like there are there are a lot of things to like, you know, to actually get there. But if you work backwards for like a few years from now, like I I think that that would be a more natural market state. But you're gonna need new payment infrastructure and solutions and products that can enable both the discovery, you know, the orchestration, you know, the payment,

Gabe Tramble (30:26) Mm-hmm. Yeah. Yeah, yeah.

Cuy Sheffield (30:49) You know, the authorization, like the the estimate, like there are all these things that need to happen for that like open ecosystem to be able to emerge.

Gabe Tramble (30:57) Yeah, yeah, I think I think that's really well said. One one thing that comes to my mind as you're saying that is like that piece right there, the just just the agent to put in like a one-liner agent paying for its own inference just opens like a science fiction, maybe not science fiction, probably not science fiction, like ⁓ thoughts about agents just running around autonomously and and self-improving. I know Anthropic just published a report, paper yesterday about the agents improving themselves and and being able to get to the point of of ⁓ continuous improvement and and development. So yeah, I just wanna to f to flag that does ⁓ creep me out a little bit.

Cuy Sheffield (31:35) Yeah. But I I think it it kind of departs from the crypto like there's a crypto world and like I I think the crypto AI world and like Bit Tensor and like it's like a fascinating ecosystem. And it like feels like it's like science fiction. Like they're like but I I don't understand you know most of it. ⁓ but I think there's this idea of like self sovereign agents that you're gonna have an agent, you know, that is gonna pay as it goes for its own inference and then it's gonna have to like create value or else it's not gonna be able to pay for everything. It's like That that's the like vision people paint. I think that is a very scary future. And that's one of the reasons why it's like we think agents should be working for an accountable, you know, to a developer or an individual. And you should be able to tie those two together. And if you're gonna interact with an agent, you you you should probably wanna know, you know, who that agent is actually working for instead of that agent that's just working for itself, you know, it's self-improving towards towards some goal. And so it's kind of like that my pet peeve is anytime someone says, Yeah.

Gabe Tramble (32:07) Yeah.

Cuy Sheffield (32:35) Agents are gonna have to use crypto and stable coins because they're not gonna be able to go and get bank accounts. And I think that like the truth is that well, in almost all cases, an agent is created and works for an individual who's a consumer or business. And that consumer or business can get a bank account and already has a bank account and a card. And so you it's just about you know delegating and giving those credentials to the agent like you would an employee. I think the only case where that's not the case is that the agent is self-sovereign. Is created and kind of let go into the wild and they're not working for anyone. And in that scenario, that's terrifying. And like I think that it's very hard to be like, yeah, let's let's go and like support this agent that you know reports to nobody and like is just out there in the wild and let's sell it a bunch of things when we don't know what its objectives are. Like I I think most merchants would look at that and say, That that seems like that's pretty high risk. Like I'd rather sell to an agent that I know is. being operated and controlled by someone who has a bank account and a card that, you know, there's some accountability on.

Gabe Tramble (33:38) Yeah, yeah. Yeah, I I appreciate you ⁓ posing it that way. 'Cause I I think like for navigating this space, everyone has like their mental models and you might get stuck in a mental model either from like reading stuff or just kind of you know information just pounding you on the head all day and you can get stuck in those those ⁓ like ideologies and then you just literally cannot see the forest right ⁓ in it at the end of the day. So ⁓ it just reminds me also I when when I was growing up, my parents got me like a kid's account at at the bank and that was my first account. And I had that like kids account, bank account, debit card, but it was tied to their account. And once I hit 18, then ⁓ it graduated me to my own account, right? So I can also see that happening. And and we we talked to Work OS, and they have a similar structure for doing credentialing and and ⁓ authentication for agents on websites. They they launched a auth.md. And yeah, they have the same kind of mindset where they allow the agent to maybe like anonymously interact with the website, but then later.

Cuy Sheffield (34:37) Mm-hmm. Yeah.

Gabe Tramble (34:48) the website might require the agent to like bring a human or an email or something like that. So I'm kind of seeing like ⁓ pattern parallels between ⁓ you know, on the on the website authentication side and and and on the payment side for for how you're looking at the space.

Cuy Sheffield (35:04) Yeah, I I think that the the analogy is almost like just from like a very, very simple sense of every individual can now be their own corporation. Where like corporation, you know, has yeah, a set of employees, and then you know, those employees are delegated certain responsibilities and you know, certain budgets and certain amounts of money, and those employees can go do things, but like they're accountable back to you know the corporation. ⁓ now that's like every consumer can have you as many agents as they want working for them, doing different things with different roles and responsibilities. And and I think to me that's awesome. Like that's what gets me really excited is like, you know, for you know a pretty small accessible amount, depending upon like what the agents are doing, being able to have your own fleet of agents that are working for you and like, you know, doing different productive tasks. ⁓ and so I I don't think it's like that different from that idea of you know every individual being a company.

Gabe Tramble (35:53) Yeah. Yeah. Cool. And and you you touched on ⁓ the tap or the the trusted agent protocol. ⁓ can you and and the trusted agent protocol is like a a subsidiary under your ⁓ visa intelligence commerce like platform, right? Which is like agent agents and commerce. Can can you explain the ⁓ the VIC or or Visa Intelligence Commerce stack and and then where like tap fits in there?

Cuy Sheffield (36:30) Yeah, I I think at a high level, like Visa Intelligent Commerce, you know, enables agents to be able to have access to a card that is tokenized, that's secure, that can have certain controls and and permissions, can be identified as an agent. And so that the payment functionality is pretty straightforward that we're able to solve, ⁓ using a lot of the infrastructure that Visa's kind of built out in tokenization and some of the same technology that powers Apple Pay. you can now you know be able to use as an agent that someone puts a card on file, you tokenize the card so you have a secure you know representation of it that an agent can have access to with certain you know guardrails and bounds around it. So I think that was like the first thing. Like, okay, if if an agent's gonna have have a card, you shouldn't just give it a 16 digit pan. Like, you know, there should be a tokenized representation of the card that is more secure that the agent can have. So I think that's the first thing. I think then the question is, okay, well, like That agent needs to be able to take that card and go somewhere and spend. And I think the the very real challenge that we saw was that a lot of merchants, particularly like when we started this a year ago, it's a lot of merchants just block bots. Like they're like bots are bad. Like they they've like for the entirety of the internet, it's been if a bot shows up at a website, like, you know, it's there doing something you probably don't want it to do. ⁓ and so I think that merchants to it's it's one thing to just give an agent a card, but like, How do you help that agent kind of be able to get to a checkout and then be able to spend on the checkout? And I think that that's that's kind of the gap and the barrier in the space right now. And part of it is that merchants and through providers, you like Cloudflare and CDNs, they've all these things in place that are specifically set up to block bots. How do you differentiate between a good bot and a bad bot? And so what we're doing is kind of building this identity system in tap where If you have an agent that has a card that has been tokenized through Visa Intelligent Commerce, being able to then have a signature that that agent can carry. And so then when they show up at a website, being able to say that this is a trusted agent that that merchant should let in versus this is a bot that's there, you know, trying to do something malicious. So I think that's the high-level concept. I think we're making a ton of progress on it, but it's it's still incredibly early for most merchants. ⁓ and I think that. Right now, you know, the the bigger challenge is how do agents discover inventory ⁓ and be able to like you know have merchants identify good agents, bad agents, and let them get to the checkout. If an agent can get to the checkout, like now agents can have a tokenized card. If they can get to the checkout, like they can pay, but it's that like in-between layer. and so we think there's a big role for for tap to play. We think there's a big role for UCP.

Gabe Tramble (39:16) Yeah.

Cuy Sheffield (39:21) I think there are gonna be a number of kind of protocols and standards that have to exist to enable like a headless checkout that also has trust and identity built into it. ⁓ so that's like the core stack right now. And then you know what I've been focused on is is building out Visa CLI, where Visa CLI is is very much, you know, rather than focused on you know traditional kind of e-commerce you know agentic use cases, Visa CLI is like a very developer vibe coder native.

Gabe Tramble (39:38) Yeah.

Cuy Sheffield (39:50) If you're inside a coding platform, you know, how can you put a card on file and then be able to buy inference, buy image generation, buy data? ⁓ and so Visa CLI is is kind of closer to the X42 space where we're actually trying to enable a card-based experience to pay for X402 endpoints. I think in the future, you know, all these pieces will come together. And what I'd like to see is you could have an agent, you know, wherever that agent is, they can put a card on file. Yeah, they can get a trusted kind of signature that they're a good agent through tap. They could then spend any traditional e-commerce checkout ⁓ over maybe a protocol like UCP, or they could go buy inference and buy data. And I think you'll see different use cases where, like, you want to book travel. Well, maybe you want to buy some weather data first, ⁓ and you want to do some like small microtransaction, and then that ultimately leads to completing a larger transaction. ⁓ and so I think that you kind of need to have a set of

Gabe Tramble (40:48) Mm-hmm.

Cuy Sheffield (40:50) primitives in infrastructure that can support across the board everything from developer facing micropayments all the way up to like executing a traditional e-commerce payment, you know, at the point of sale, you know, with an agent. So we're trying to just build that kind of core layer across the stack.

Gabe Tramble (41:06) God. Yeah. Yeah. It feels it feels to me like now, and it's still super early, just even on the payment side, and and there's like all these things that need to happen. But for the people who are like running at full speed and and at the edge, it seems like the real break is the the distance between the agent and traversing the website to get to the checkout. That just seems like There's a few companies like Browser Use or Browser Base that are like actually just using the human ⁓ browser interface to get to the checkouts. Can you kind of explain a little bit more if you think that, in my experience, that's kind of the big gap right now. ⁓ more on your end of like where you think that's gonna tighten up, or even if you think that is a gap in the first place.

Cuy Sheffield (41:49) Yeah, I I definitely think that that's that's a gap. ⁓ I think that like most of agentic commerce, you know, experiments and pilots and kind of early versions of it has been requiring an agent to navigate you know a website like a human. ⁓ I'm interested to see how browser use and like computer use and like how how all of that improves. But it doesn't seem like it has gotten to the point today that works well enough, you know, efficiently enough at a low cost. ⁓ to be able to actually make that a mainstream use case. So I don't know, like it could change, it could get much better. But like today that hasn't, you know, been kind of the sole solution that you know everyone said, great, we're we could just use browser use and get to any checkout that we want. And there are a bunch of nuances of the way checkouts work at many different merchants, that you run into issues when you just try and do browser use. I think having purpose built protocols, yeah, like UCP, yeah, make more sense of like a headless checkout that is API based that you don't have to send a browser ⁓ an agent you know over the browser. ⁓ and so I I think that you'll see like a combination of both. ⁓ but I I see a need for like yeah these specific protocols that are built for a merchant to be able to communicate you know with an agent without the agent having to to go over the site. And I'm excited to see merchants, more merchants start to onboard into those protocols, you know, start to learn how to use them. ⁓ and I don't know how far away we are, but I I think that there's there's a good recognition across the industry that you know merchants that want to sell to agents are going to have to do some work. ⁓ and you know many times their platform like Shopify has enabled and kind of does ⁓ makes it very easy to do. They're gonna have to represent their inventory, make it easy for agents to discover, and then be able to execute payments through those protocols. And and that's probably a more sustainable long-term solution than than browser. ⁓ browser automation, but we'll see. I I don't think anybody knows. Like it's great to have people trying a a bunch of approaches.

Gabe Tramble (43:49) Yeah. It's it's gonna be interesting with stuff like Shopify or these these platforms because you're probably gonna have like real world market makers, you know, doing like alg algorithmic trading for goods. And yeah, I'm just I'm curious to see like what emerges as ⁓ as it's easier and and and you you can like machine interact with all these different platforms, machine to machine. So that that's something that yeah

Cuy Sheffield (44:15) I I think that's that's one of the reasons that merchants are nervous, that they're like, Yeah, if are we just gonna put all of our inventory into this protocol and now we just become a catalog as an API and we lose our brand, we lose our value, like it's it's a scary thing, you know, to give up, which is partially why I I think there's a big opportunity for new merchants. ⁓ and I think like more often people when they think about agentic commerce, ⁓ they think about existing merchants, you know. making you know products available you know to agents. And that's to me almost like it's like we the internet's coming around and you know how are existing brick and mortar retailers going to adapt and you know making products yeah available to to users on the internet. ⁓ and I think that it usually starts with like some success cases of brand new merchants, brand new retailers, brand new companies that are just built on day one, their entire business model, you know, their entire mission and vision is selling to agents. And I think they're the ones that have the incentive to go and try and figure this out. And then you have some breakthrough successes. And then, you know, what they do gets adapted. And that, you know, the traditional merchants that have, you know, big e-commerce franchises today, you know, they start to open up over time. But it's just so hard for me to imagine that it's going to be an existing merchant that has a great business that they like in a format that they understand that is going to be on the bleeding edge in terms of experimentation around how to sell to agents versus. just a new merchant class that's emerging. And I think the barrier to entry to building a merchant is just getting lower and lower. And I think that's one of the amazing things about, you know, vibe coding. It's like, you know, anyone in an hour, you know, can build something that they could then sell online. ⁓ and so I think that there'll just be many new merchants that are created. And I think some of those merchants will take an approach to say, we're going to build exclusively for agents. Maybe there's no way a human can even pay at them. Maybe they don't even have a website. Like it's just an endpoint, but we're to build for agents.

Gabe Tramble (45:50) Yeah.

Cuy Sheffield (46:10) And I think that once you start to get some success cases there, then more traditional merchants will say, okay, let's let's dive in. But it's it's just hard for that to to start with a traditional merchant.

Gabe Tramble (46:10) Mm-hmm. Yeah, that's super interesting. It reminds me quite a bit of like Ghost Kitchens. ⁓ and and just you know, this new entrance of player. I d I don't know if Ghost Kitchens are a success case, but very similar ⁓ kind of parallel there.

Cuy Sheffield (46:38) For sure. Apologize. I I gotta get I gotta get good. We we could talk for hours. I I I gotta get going if you wanna if you wanna wrap up. I don't know the the best way you

Gabe Tramble (46:42) Yeah, yeah. perfect. Okay. Yeah. Yeah. Yeah. No worries. Yeah. I was I was watching the time too. ⁓ I'm gonna do one last question and then just like an alpha, like yeah, any like what alpha do you have and then and then we can wrap on that. Yeah. Well let me let me let me let me put it in the camera. So yeah, Kai, thanks. Kai, all the stuff that you've been saying so far ⁓ has has been kind of putting this this model and and giving us this example of what to look out for.

Cuy Sheffield (46:58) I think my

Gabe Tramble (47:12) In the ecosystem right now, where do you think as we wrap here, where do you think is the last place where alpha lives? And, you know, if you were to a new builder, where would you focus on right now?

Cuy Sheffield (47:24) Yeah, I I think it's it's pretty simple advice yeah that I I give to everyone that I talk to now is like the best way to understand agentic payments and where it's going is to just start making agentic payments and just start using these products. And like I've been trying to buy something, you know, with an agent every day. And just like, what are interesting things that I could go out and try and buy? How does it work? Where does it break? and I think that's that's where the alpha is. And like there's so many people who you know just want to talk about agentic. And then you ask them, like, have you ever bought anything with an agent? Like, are you are you using any agents? You're right. They're like, No. It's like it's just like you can talk forever, but it's like you have to just use the products. And as you use the products, you figure out what works and what doesn't, and and then you iterate from there. So that's that's my advice.

Gabe Tramble (48:10) Cool. Well, cool, man. Thanks for sharing. And yeah, I'd love to have you back on soon. Yeah. See ya.

Cuy Sheffield (48:15) Thanks for having me.