Getting Started with Agentic Payments on Crossmint

In the earliest innings, everyone sort of kind of assumed, yeah, at some point agents will probably pay for things. But as soon as Muse and Instinct came out, like really back-to-back, and even in that first week, one of the first things that Muse suggested, gave you three things. It was like, let me check your email, let me do X, let me buy you something. It was like, oh my god, here's an inflection point. And we went from like, agents will probably buy something one day, to, oh of course agents are going to shop for us.

GM, GM! Welcome to the show, Alex. I'm ecstatic to have you on today, man.

Thank you for having me. It's always so exciting to talk about what we're building here at Crossmint, and I just genuinely love telling more people about what the future is going to look like.

Absolutely, absolutely. And yeah, today's topic and what you guys are building, it's really an emerging space right now. Everyone is really hot on what models are coming out every week, but I think we're at a point where we have enough intelligence that people can start building things on top of it, and payments are a big one.

It's funny you put it that way. I actually was at a dinner with a bunch of AI folks last night, and there were a lot of us kind of sitting at one part of the table saying: do we need these things to get that much smarter, or are we still just like, we still have to just mine deeper to figure out what are all the ways that we can put this incredible intelligence that we already have to optimal work? So yeah, I think that the tools that are already in place are incredibly capable and we're ready to build some really cool things with it.

Absolutely, absolutely. And I've seen my fair share of that as well. But before we get into all that great stuff, would you just like to give an introduction about yourself?

Absolutely. So I've been in the tech world now for almost 15 years. I worked at a startup accelerator. I moved on to work at an IoT company building hardware and software to remotely monitor telecom sites. And it was actually while I was at that company in 2016 that I read some of the early work thinking about how Ethereum and smart contracts could be used to enable machine-to-machine payments. And that's what kind of sent me down my Web3 journey. Back then, like I'd known about Bitcoin for a long time and was interested in it, but the idea that we could make money more programmable and let machines control money was one of the original things that got me interested in this space. I went on to work at Chainalysis for almost 3 years, but then left Chainalysis to start my own company. I built a protocol called DIMON, where we were putting machine identity onchain and had the dream of eventually getting into machine payments as well, although we were solving a bunch of other problems around data management and data verification. And after 5 years of building DIMON, I left DIMON to come work at Crossmint and lead agent payments here at Crossmint. So it's been a journey that has taken me in and out of the Web3 world, it's taken me in and out of payments, it's taken me kind of like in and out of a few of these different areas, but somehow I always find myself intersecting with how is money getting more and more programmable and how are non-human actors going to move money.

Absolutely. That's quite the amazing journey and I think you landed in a great spot, as I believe this intersection of blockchains and AI is a natural fit. We haven't been able to get humans quite onboard with using wallets and private keys, but all these things are kind of very well set up for AI to be able to understand and work with. It is what it is, but yeah. I suppose from that standpoint, we know a lot of agentic products are heavily developer-oriented, and that's just the near-term opportunity and what people are paying for now. But a lot of people don't realize that agents will also eventually start to spend online and kind of be extensions of ourselves. For the person watching today that doesn't quite realize that, did you want to give an overview of what your perspective is on that, and whether or not we should be giving agents our credit card to spend on our behalf?

Yeah, so I'll talk a little bit about AI as a journey, because I think the moment that took us from thinking about LLMs as a chatbot to really thinking about agents was of course OpenClaw late last year. And I remember getting into OpenClaw and starting to build my own agent. And I'm a relatively tech-forward person, but man, I was like, I'm not a dev, I have no interest in DevOps! And I started creating my own skills, and then you start seeing other skill libraries, and it didn't take me very long for me to just say: I would love for somebody to make me an agent with a thousand skills built into it that are already really good, and just come along and give that to me. And I've used everything; I used Hermes and that was like a step change. But I think that obviously the new wave of consumer assistants is finally hitting that level where maybe it's this version, maybe it's one more iteration, but we're getting closer and closer to actually crossing the chasm. And it's really funny, here at Crossmint we've seen some data. Many of the early OpenClaw people were iterating on the edge and were playing fast and loose, and basically said: sure, I'll give my OpenClaw access to my credit card and I'll just save the card number someplace that I think is relatively secure. There was quite a bit of credit card leakage in that world. That is maybe fine if you want to take on that risk as an individual, but if you're trying to offer an agentic service to somebody else, there's a reason we have compliance and regulations around how you should handle those card numbers.

So in the earliest innings, everyone sort of assumed, yeah, at some point agents will probably pay for things. But as soon as Muse and Instinct came out, like really back-to-back, and even in that first week, one of the first things that Muse suggested gave you three things. It was like: let me check your email, let me do X, let me buy you something. It was like: oh my god, here's an inflection point. And we went from like: agents will probably buy something one day, to: oh, of course agents are going to shop for us. And the last high-level observation I'll give just to get the ball rolling here is, I see a lot of people talk about what is the thing that agents are going to buy, and there's a lot of debating and arguing on Twitter: is it going to be flights, is it going to be groceries, is it going to be this or that? In reality, I've talked to hundreds of people about this and people are all over the place. Like for me personally, I'm a Delta loyalist, I don't think booking flights is that hard or complicated, I have some nuances, I change the way I want to buy on a few factors, and it's just easier for me to go book that myself. Just easier. Where I have found agentic commerce to be useful is on things that I periodically reorder, but don't make sense to put on a subscription. And so my agents know a list of things that I buy and I can say: reorder the dish soap and the multivitamins, and they're just like: okay cool, I know what you want, I'll go to Amazon and get it for you. And I think if you talk to people, they're going to give you a range of answers. There's no right answer to what agents will buy first. And that, of course, is what opens up the opportunity for many builders and for thousands of companies to do this successfully.

I absolutely agree. I think there's an ocean of opportunity. It's just a matter of figuring out what people want and how they're going to go about using AI in their day-to-day lives. Developers are one thing, right? We kind of have our daily habits of things we want to do and how we like to use our agents. But for the average consumer, those habits are going to be different, right? It's like maybe somebody just doesn't like standing in lines, and they just want their agent to go buy them groceries from Instacart or something like that. Or it could be just the little things that you want to save time on. Yeah, it's going to be very interesting to kind of see it play out. But I know Crossmint is kind of at the forefront trying to prepare for that eventuality to get agents being able to buy things for us. Did you want to give an overview of what Crossmint is overall, and then we can get into the agent stuff after that?

I'd love to. So Crossmint has been around for about 5 years, and the company began as a really easy and seamless way to run an NFT launch. One of the original innovations of the team was to make the bridge from fiat to NFT purchase a lot easier and faster. And so they were really the first team to create a headless checkout where users could show up with a credit card or Apple Pay and purchase an NFT. That eventually, you know, NFTs had their cycle and are not quite as prominent of a commerce category as they once were, but of course memecoins have become quite prominent, and then the largest growth area, of course, for really any kind of crypto activity today is in stablecoins. And so over the last few years, Crossmint has become a stablecoin company. We facilitate onramps and offramps to some of the largest companies in the world as well as wallets, so MoneyGram, Robinhood, and a bunch of other fintech companies. About two years ago, actually coming up in November will be two years, quite a bit before I joined, the company published its first foray into agentic commerce, which was called the Great Onchain Agentic Toolkit, GOAT. It was an SDK. And the premise behind GOAT actually still lives in our products today, which is: how do you give an agent a wallet in a controlled and in a permissioned manner? Crossmint, as a wallet provider, builds smart contract wallets. So we actually encode all of the policies that are relevant to controlling those wallets onchain, and we use delegated signers. We actually have multiple signing layers: we have operational signers and then recovery signers. But basically, it's pretty simple. If you want to give your agent access to a wallet, provision a wallet to the person and give the agent a delegated signer. And in fact, you might give it multiple signers. Maybe it has different budgets, it has different purposes for each signer, and maybe you have different agents that each have their own signers. It's a very flexible construction. And so we were able to put this out almost two years ago as a way for you to equip your ChatGPT with a wallet. Since then, the agent team has iterated and evolved quite a bit. We still support—and we'll get into this in some demos and some other activities—we still support wallets for agents, including onramps of course, so you get all that in a nice, clean package if you need your agents to transact in stablecoins. And we've added cards to the mix. So we support some of the latest agentic commerce protocols from Visa and Mastercard. One of the really interesting things that we're seeing is virtually every payments company on Earth is now starting to unveil, well, they're planning to unveil, some way for agents to access those funds. So all of the big payments names that you could imagine are going to say: you have an agent, here's how your agent can use your, fill-in-the-blank, account: PayPal, Cash App, whatever. The last part of what we've been building on the agent team is a checkout API, because once you've given the agent access to some money, it still has trouble buying things. And the world of checkouts—we can talk about this more in a bit—the world of how your agent actually goes and executes a purchase with a merchant is not a very well-paved pathway today, and it's one that is set up to evolve quite a bit over the next few years.

Awesome. That's all exciting stuff to me, especially you guys announced the agent commerce toolkit yesterday, I believe. Do you want to tease what that is, and then we can get into the screen-share demo?

Yeah. I would say it's both a combination of the tech itself and sort of the user journey. The products themselves are the agent cards that we have, the agent wallets, and the agent checkouts API. The major changes include: one, finally making all of that self-service. Being a regulated financial institution and selling regulated products like onramps and offramps, most of what we do, most of the business we do, is not self-service. We have to go through KYB, we have to spend a lot of time with our customers. Not quite as cumbersome as getting a bank account, but we have compliance obligations we have to fulfill. That's not the case with these products. Now, there are some risk checks and risk considerations, especially because we have sort of unique access to Visa Intelligent Commerce and Mastercard Agent Pay. So we have to be quite considerate about how we handle those things, but at least we can help get developers to production without throwing up a gate in the way, and that was really, really important to us as we were bringing these products to market. So now we have the full package of the SDK, we've made it all self-service, and crucially, we've put out not one, not two, but three sample apps! And we've made it as easy to fork as possible, and for somebody to get started and say: I want to build a wallet, I want to build an agent that does XYZ, or I want to build an agent wallet—here are all the different ways you can make that possible. And so that's what we're going to talk about here for the most part in the demo, showing all the different cool sample apps that we have.

This is the way. You've got to make it easy for developers to get up and going and shipping something, as there are lots of intricacies in getting this side of the equation set up. You think high-level this would be easy, but once you start, you're like, oh, hold on a minute! But yeah, I love to see this in action, so let's get into it.

Let's hit it. All right, pulling up my screen share here.

So we're going to start with our sample app. It's probably the most flexible sample app, and I think it does a really nice job of showcasing a handful of the particular flexibilities that we have. So here we are looking at this sample app, and right now it is set up to display what an agent iOS app would look like. So here we have a mobile app just to showcase, hey, one of the really important things to us is everything we do is headless. We are not bolting on—one of the most common things you see in the agent world today is many agents are using the Stripe Link wallet as a means of accessing payment methods. That's a great product, but to me, it violates the cardinal rule of building a consumer product, which is: you are now taking the consumer out of your app and you are bringing them to a third-party app. So what we've done is we've built everything to be white-labeled completely into your own product interface. So whether you're building your own agent on a mobile app, a desktop app, whether you're just creating an iMessage agent or even a WhatsApp agent, or making it a ChatGPT plugin, Crossmint has you covered in terms of how this agent can fit your brand and fit the user experience that you want to create. We have two separate apps. The one you're seeing here is our cards demo, and then we actually have a separate one that we can send out, which is our machine-to-machine payments demo. So on the one hand, we're saving a card in order to purchase things on the web, and in this demo, we're creating a wallet, topping up that wallet in order to use x402 or other resources like that.

Some of the other functions that we have: so you might have noticed I have a wallet already saved here. In my case, this is my kind of daily spender, my Amex SkyMiles account. I could add multiple cards. And one of the other really important things, of course, that even some of the folks like us really on the edge of the technology are getting used to, is the idea of: how do I control how much my agent can spend? And there are a variety of ways you can do that. The card networks themselves are starting to build some of these controls into these agent protocols. And then we actually layer in, in some cases, additional controls via our SDK where you can set a budget, you can set authorization gates, you can create additional controls for the user. That's part of what you do as the developer building on our platform: you get to decide which of these functions you want to use. Because of how we handle that authorization, it allows you to create a lot of really interesting and super flexible outcomes. So you could pre-authorize an agent to purchase something and say: hey, I want you to buy me this thing, but wait for the price to drop from $60 to $50, and then buy it right away. And that's, I think, one of the directions that people are going to go and use agents much more adaptively. I would say, and this is anecdotal, but probably 95% of the agentic purchases that happen today are people who say: I want this thing, buy it now. But we're starting to see an increasing share of: I want this thing, watch the price. And I think the holy grail, maybe the next layer beyond that, would be: okay agent, you have a budget, and I want you to figure out which of these things I need and when to order them. I don't know, maybe one day agents will get credit! There's a lot of weird things that might happen in the future. Let me take a quick pause and see if you have any questions so far.

This is really awesome. Love the demo, love how slick it is. I was going to mention, that 'wait for the price to drop', I've done this in the past. I don't know if you remember when the PS5 and Xbox came out and you couldn't find it in inventory. I personally made a Selenium script that would go and scrape the different websites. Basically, this kind of evolves into that, and now that we can kind of make it so that the agent can purchase it for us instead of just alerting us to go check it out, I think it's pretty crazy. And this was kind of the side of equipping your agent with a card, provisioning it for spend. But when it comes to actually going and doing the checkout part, the payment part, I'm really curious to kind of dig into how you guys solve for that, because that's one of the bigger pain points. What can an agent buy, that's one part of it, and the second part is a lot of checkout pages kind of block an AI from being able to enter details there. So yeah, keen to learn more on that.

Yeah, that's a great question, and we'll progress through a few of the other pages that I have queued up here. Just as a quick orientation on the docs, we have a section for agent cards, and this governs and explains how our agent card product works, how users can save cards. Technically speaking, there are different ways to save and use cards depending on whether or not that card can be tokenized by the card networks. This is a term that they have in terms of how the card gets saved, and basically the relevant point for developers is that we have secure ways to handle the card whether or not it can be tokenized, and we can fall back to different pathways that matter. When it comes to checkouts, just another note for developers out there, Crossmint itself, we don't handle the discovery layer. So we don't have a catalog. We may eventually fold that into the platform, and there's a bunch of folks that offer catalogs out there. But if you want your agent to have up-to-date information on what's in stock and what's on the store, there's other APIs you can go hit. We become relevant once the agent knows what it wants to buy. So basically, what you'll do is you would submit a URL of what you want the agent to buy, and you can actually provide also some context about that. That context is typically stuff like buyer profile, home address, work address, shipping preferences, or if we're ordering an item, the context on the size. So I actually want to buy this pair of gloves, and I'm going to say: standard shipping, use a card, XL. Now, as an agent builder, you could store all of this information. You could have user preferences saved that could have told this agent ahead of time: always use standard shipping, if you're ordering gloves, XL, just use a card. I'm specifying card because we can also access other payment methods like Shop Pay, saved cards on file, whatever it might be. So this is yet another one of our demo apps that we're going to be releasing more publicly soon, called Open Commerce. You can actually plug it into your own coding agent or your bot in general. I have it hooked up to Grok bot and it works amazing. So what happens? I've now given my agent an instruction to go buy this, and it's calling the Crossmint APIs. This is a fun demo because you get to see the actual browser automation happening in person. The way things work today is that 99.99% of the web is inaccessible via API. Basically other than travel, most of the web is not accessible via API at all. So our tooling checks for one of the agent protocols—UCP, ACP, etc.—but we didn't find that here, so we're going to move on. Now we have to handle this checkout with browser automation. We have built our own browser automation in-house. We lean on Kernel for the browser itself, but we do all of the execution and the logic and the handling in-house itself. We've layered in a few important pieces for that, like persistent sessions so if a user logs in once, they can actually expedite the checkout process the second time. And we allow you to save items to your cart, come back later; there's all kinds of really nice configurations that you can build as a developer with what we have.

Now you can see our agent is going through the process here on the browser, filling in all the information. On average, we're seeing this take about 60 seconds or so, so it's coming back pretty quickly. Now it has a quick question for me; it also comes back and will ask you if it gets stuck or doesn't know how to answer something, it will return back to you and have you fill it in. Once again, I could save my card on file or tell it my phone number and then I wouldn't even have to be human-in-the-loop on this. The final question is: where is this all going? As I mentioned earlier, very few merchants have released anything in the way of UCP or ACP integrations. The really big hurdle that's going to happen over the next few years is that merchants are going to require things like Know Your Agent identification and anti-fraud or trusted agent protocols. So when your agent shows up to a merchant in the future, it's going to need to identify itself and say: hey, I'm Alex's agent, here's why you can trust me, and here's my W3C credential and some of the progressive disclosures that I'm handing out in order for you to believe all this. We are integrating all of that into the Crossmint checkouts API so that as a developer, you don't have to think about it. The idea is that while we're handling this through browser automation today, and you can see we've handled this order here in about a minute, in the future, we are just going to orchestrate across the protocols. So the agent is going to show up and in order to buy something, it doesn't need browser automation; it needs to say: find me the UCP endpoint, let me identify myself as Alex's agent and whatever other signals the merchant needs, and here we go, let me buy. All right, I've been talking for a long time, we purchased something on the way, what questions do you have?

Absolutely. Just wanted to confirm that this was a legitimate purchase online and you're going to get these gloves, right?

Oh yeah, I'm going to get these gloves. I'll show you the receipt here. Probably came to my, yeah, let's show you the order confirmation. Got to flip windows here. There they are. And as you can see, it just came in 0 minutes ago here.

Look at that! Yeah, that was a legitimate end-to-end flow. For folks watching and not kind of realizing what just happened: an agent went and bought a pair of gloves for Alex on his behalf, made a real payment. Welcome to the future! And yes, like you mentioned, there were kind of some parts in there that were human-in-the-loop where you had to accept the actual payment and verify some things there. But yeah, eventually this is going to go in the autonomous fashion where you just have a set of things you want to buy, like multivitamins or toothpaste or whatever, you set it and forget it, and it's on its way. What in your mind is kind of the blocker or inhibitor of merchants activating that layer? Can you point it to any particular thing why some people are slow to move?

I think there will be a lot of people who would say, oh, this protocol is not fully fleshed out and there's some things that aren't thought through and we could point out technical holes. I think it's purely behavioral. And I think that there are, especially larger merchants, they can wait. They can afford to wait. One of the stats I point to is only 2.5% or something like that of US households have a subscription to an agent or to an LLM. And we are still very early in the adoption curve of these tools just generally. So even though you and I are on the technical cutting edge and love to live and breathe this stuff, it will be a while before the critical mass is really begging for this.

Absolutely, absolutely. It's just a matter of time and these merchants need to realize that AI agents are going to be an emerging customer base of theirs. Maybe not today, but definitely tomorrow. So being able to kind of prepare for that and build for that, I think it's vitally important regardless of the size. So definitely if this intersection of payments and AI is something that interests you, all those resources for you to get started will be in the description of the podcast below. And with that, Alex, just want to thank you so much for taking the time out of your busy day to come chat with us. I just want to wish everybody a very happy Friday, happy weekend wherever you may be, and we will catch you here next week for another great episode of DevNTell. Till then, have a good one, folks. Cheers!