Colin Banville, Stripe | How Payments Change When Agents Pay

Hey folks, how we doing? Hopefully you're enjoying the conference so far. Two days worth. I didn't see yesterday, but I'm sure it was amazing. We're going to change gears a little bit and talk about AI agents, payments, including stablecoins.

Um I'm going to cover So first of all, Colm Banville. So I work in our AI solution architecture team at Stripe. I also happen to lead the agentic commerce side of things, which includes um the stablecoin um payments. We're going to talk about what changes when agents pay today move into machine payments, which is relatively relatively new, a little bit more advanced than it was a week ago, and just some key takeaways. So let's have a look at the history.

Um probably a lot of you don't know that Stripe actually first got involved in crypto back way back in 2014. It's a long time ago now. Um there was Bitcoin acceptance at that point. Did that for a number of years, then quietly removed themselves from the crypto space for various reasons. And fast forward to 2021, we feed acquiring that launched.

So that was for crypto exchanges and marketplaces. 2022 was crypto on-ramping and stablecoin payments launched. Into 24, where we have um stablecoin payment launches um themselves. And 2025 was the acquisition year. That's where things really um came together for Stripe, shall we say, from a crypto stablecoin perspective.

We acquired Privy, and we acquired Bridge. And if you haven't spoke to the Bridge guys, I believe they're in the networking area, so make sure you have a chat with them. If you think of Stripe, it has It's an infrastructure for financial services, yes, but from a fiat perspective. Um what we've built through these acquisitions and Stripe itself um is fiat rails over here, but then you have crypto and stablecoin infrastructure over here. So, a complete solution for that.

Base layer being how the money moves, blockchains obviously. I'll talk about Tempo, which is our layer one payments blockchain in a little bit. We then have Privy, which is our wallet infrastructure, digital asset management, bridge, so all of the stablecoin money movement cross border, all of that really fun stuff. And then sitting on top of that is Stripe itself in terms of making sure it's agnostic to the payment infrastructure and the rails used. But, what we're seeing is with the explosion of AI, I'm sure you've already heard about it at the conference.

Um it's reshaping paths to purchase. Okay, so the behavior of users is changing, all our all our behavior is changing. Like, does anybody use Google search anymore? Probably not. It's all in chat to do your research.

So, AI being the new search is one area where you can now go agentically and actually complete a purchase inside of ChatGPT, inside of Gemini. You've also got AI powered advisors. So, those advisors sit on websites, for example, giving merchants more brand control, um but also it can be in the search bar, it could be a pop-up agent that will do discovery and complete the the purchase for you all within um your own website. So, no third-party involvement there. And thirdly is multi-party agent interactions and that's we're going to focus a little bit more on that today.

So, that's machine-to-machine payments and how we can facilitate that. So, if you think about and I heard the panel beforehand talking about tapping your card and needing to change um the kernel for the point-of-sale systems, right? Well, if you look at what changes or what happens when agents pay, right? A typical when you tap your card, a typical transaction looks like user to merchant processor payment rail fiat issuing bank, okay? When we introduce agents, we're introducing two new agents here.

There'll be some type of discovery agents, there'll be some type of payment agent. Okay? So, a user will give the agent permission to at make a purchase or act on their behalf. Okay? The agent then goes, checks out the product that needs to be purchased, um it defines the intent um to make that purchase um for the on behalf of the user.

It then at that point in time is just passed to the issuer bank, right? And the issuer bank then does a fraud check, at risk scoring, funds are reserved. Um once the funds are reserved, the merchant has to then go and approve. Once the merchant approves, the agent actually completes the purchase, right? This is all happening in the background when you go and tap your card.

It's an extra layer of complexity when there's an agent involved. Okay? So, or multiple agents. And that complexity is more around human commerce or I'm calling this legacy here, but basically manual checkout. We've all hit the capture um on the websites where you've seen put the code.

Agents themselves can like can reason very and complete complex um integrate like complex challenges. But when they when they go to make a payment, that's where everything kind of falls apart, right? So, it's high friction with like captures, batch settlement, which can take days as well. And it's it's GUI driven, right? So, it's for using a mouse, it's for human human and eyeballs essentially.

The new way of looking at that is we need programmatic execution. We need zero friction, right? And we need instant settlement, and it needs to be API driven. Okay? And this is what um Stripe is attempting to solve for um by releasing only last week our machine payments protocol.

So, what is the machines payments protocol? Um it's two it's two layers. Okay, so base layer being Tempo. This is a layer one um layer one blockchain that is co-developed with Stripe and Tempo. And specific to payments, right?

So, it provides instant settlement, high throughput. Okay. Sitting on top of that is the machine payments protocol that we released last week. This allows agents to discover, negotiate, and actually complete a transaction. We released it 1 week ago.

We've had over 900 agents in the space of that week complete over 34,000 transactions. These are microtransactions, to point out, but that's significant. I think within the week, that's the same number that the XFP protocol has hit to date. So, what actually happens if I step you through a machine payments or agent-to-agent transaction? First of all, the agent programmatically requests a resource from a service or to a service.

There's a challenge there. The challenge is, "Okay, right, this is the price. The These are the payments payment methods that are that are accepted." The agent then goes and verifies and delivers, or sorry, instantly authorizes the payment, then verifies the um resource and actually delivers it instantaneously instantaneously. How it does that is underneath the machine payments protocol, we have a a sessions primitive.

So, if you're building on this, you don't want to have to go and obviously approve every single transaction. Microtransactions could be fraction of a cent. Um the sessions primitive unlocks that. So, you set up for any particular session a predefined amount. And so, you could set up, I can spend up to 200 $200 um within this context for this session.

Okay? That unlocks streaming payments um with with primitive, and you can go ahead um and actually, you know, make multiple transactions. Could be 1,000 transactions up to $200 in that specific um session. And it's built so that it's the payment infrastructure that's used is dependent on what the transaction needs. Okay, so that could be stablecoins, right?

So, USDC um is supported. From a stablecoin perspective, it can be your legacy card rails powered by Stripe, Visa, MasterCard, whatever um or Bitcoin lightning as well is supported for the instant settlement component. It's zero friction for developers. So, if you're an existing Stripe user, it's a very simple setup. Um you'll see here um creating a payment intent.

This is a Stripe API for actually processing payments. You change your payment method type to be MPP, right? So, it's very very simple change to actually go and start using the machine payments protocol. Obviously, for developers as well, full documentation's on our website to set up, but the idea behind this um is that it reduces the barrier to entry here, right? So, it's very very simple to to get going on this.

The protocols themselves, I mentioned X402. I'm sure many people here quite familiar with that. The difference is, I would say, is the X402 protocol was definitely foundational in terms of like what it was trying to achieve and what it is trying to achieve, crypto um specific. Um it's is exact the same schema as well. The machines payments protocol has tried to build on top of that while fully supporting X402.

It's important to point out that um we're agnostic to the different payment infrastructures that I just mentioned there. Um it's using Tempo Tempo streaming's uh streaming service for the instantaneous settlement and it's got built-in expiration, tampering mitigations, and all of that. So, sub-100 millisecond latency micro cents. Use cases. There's a number of different use cases that are currently live.

I mentioned the number of transactions we've had in the week. AI inference being one of them. That's a per token large language model billing use case. Digital content. So, obviously anything behind a publisher paywall.

You In the future, you want your agent to act on your behalf. Maybe you don't want to subscribe to that whole article. The agent can go and actually make a micro transaction of fraction of a cent, which may just pull out two sentences within that digital content, right? Whatever it's relevant to the task at hand. Autonomous browsing.

Browser base being an example here. So, you want to pay per minute for headless compute. You can do that with the machine payments protocol. Real-time search. Um you've got uh Tavelling, Exa.

So, within the same session, you can make a request and actually a payment um for that for the search APIs um in those examples. So, the idea the messaging here is that it's actually real in life today in terms of these use cases that are out in the wild, which is pretty cool. It's scary, but cool at the same time. And the key takeaway is it that it's a new financial layer for the internet, right? So, instant and scalable from a Tempo blockchain layer one payment specific blockchain.

And I I should point out again, it's agent native as well. So, MVP sitting on top of that blockchain. Um it removes all of the friction and that you see with a typical um purchase cycle um from a human perspective. And I guess the the most exciting thing is, as I mentioned with the use cases, um is it's it's actually available now, and you can go ahead and you can build on top of this as developers. And with that, I will say thank you and open up to any potential questions.

I see one hand up over there, which is great because no questions is not necessarily a good thing. I'm so sorry, I I missed the beginning, but um I was curious how many how many merchants like what how's the adoption look in the early days of of the product? Yeah, so um it's only been released 1 week. We're talking about machine payments protocol here specifically,

right? Yeah, yeah. Yeah, as of last week um I mentioned the stat around just over 900 agents, right? And processing over 34,000 transactions. The biggest of those transactions were $35.

All of the rest of the transactions, you're talking about microtransactions. So, less than a dollar. Got you. Okay. Question down here.

I think you mentioned something about a discovery layer at the beginning of the talk. How are you guys thinking about discovery of services like the agents need to find these services to pay for? What does that layer look like? Yeah, and so I I think it's discovery agent I was talking about. Um in in that instance, the discovery agent would be like, yeah, needs to find exactly what it's looking for.

I think from the use cases that I've been seeing so far, it's using the likes of a browser base. So, for your autonomous browsing, you don't need like to set up for a browser set up for a browser base account or anything like that. Um but, you're just charging for the per minute usage of that browser base. Allows the agent with a predefined spec spec in terms of what you're looking for to go off and be able to find that information. From the agentic perspective and and that payments flow when I'm talking about discovery agent to payment agent, that's definitely more like a retail focused um solution, right?

So, product SKUs specifically. But, for the machine payments protocol, obviously looks a little bit different. It will sit in the likes of these new browser-based type solutions that are popping up as it is right now. Hopefully, that answers your question. You want one or two more questions up there?

One second. Sorry. So, do you imagine your the MPP being also deployed not only in payments, perhaps in the real economy, but also in trading, DeFi, vaults? How do you imagine that agentic interaction in your blockchain? Yeah, good question.

I think the short answer to that is absolutely. Are we anywhere close to it just yet? I would say no. Right? Again, this is a new primitive that's only released for over a week.

As the use cases come up, and this is the same thing that happened we released the agentic commerce protocol the end of last year with Open AI. Same thing. It's use case dependent. We start with retail, we move across industry. And so, absolutely in terms of like trading, DeFi, potentially, there's a solution there, but I don't know what that looks like right now.

But, yeah, absolutely possible.