What 500,000 Settled Agent Payments Taught Us | Marcelo Kunze at Pragma Lisbon 2026

Hello everyone. So, in 1900, there were about 3,000 cars in France. Two brothers who sold tires decided to create a guidebook with useful information for travelers. There were maps with all the roads, hotels, restaurants, and places of interest so that people would travel more. Their goal was not to create a system of ratings for food establishments.

Their goal was to sell more tires. And they realized that it was n't the infrastructure that was lacking. There was a lack of opportunities to discover new places and trust. I believe that in the field of agency payments we are now in 1900, but the solution is not to create a guide. Let me explain, but first a little context.

As Kartik said, it's great to be here in Lisbon on this stage, because we have a long history with EthGlobal and have been thinking about agency payments for some time. It all started at EthGlobal in Prague. The project was called MCP, and it was a solution for monetizing Model Context Protocols (MCPs), with agent payments in mind. Essentially, you have an API, you make it compatible with MCP, and you monetize it with our SDK so agents can use it and pay per use. This project became a finalist at ETHGlobal in Prague.

And then, with the support of ETHGlobal, we decided to go all in decided to go all in . In Delhi, we participated in Spotlight, where we had a booth and presented the project to hackers and event participants. Then in Buenos Aires we held a Demo Day, where we also had a stand, but this time we presented the project to investors. And it worked, we got into the Colosseum accelerator in early 2026. And then we also rebranded to Frames.

So, it is the beginning of 2026. Agents have become smarter. They can reason. We have skill.md files where you can actually teach an agent teach an agent anything.

But there was still no simple solution for providing agents with wallets. A designed specifically for agents. And that's exactly what we created earlier this year. Very creatively, we called it an "agent wallet." And the coolest thing is that it came with a curated registry of tools for these agents to use.

So, you provide them with a wallet and tools on which they can spend those funds. What exactly are the tools? Anything from data providers data providers like Twitter APIs, lead enrichment, image and video generation tools, to trading tools, high-speed trading—there are a huge number of endpoints number of endpoints . It worked on a query basis: you just copied one prompt into one prompt into any client like ChatGPT or Claude, and we took care of the rest . So, server wallets, passphrases wallets, passphrases —we completely hid this complexity from the user.

At that time, it was something completely new. Nothing like it simply existed, and it instantly gained popularity. So since then we have processed over 500,000 transactions. What have these numbers taught us? What did all these transactions show?

That's what I want to tell you about. The first thing that became completely clear was what people were using this technology for. So what did the agents do with their new access to wallets and tools? Over 90 tools? Over 90 % of all transactions were for the purchase of data.

As I said, it could be, for example, the Twitter API, lead enrichment, or something like brand awareness tracking . That is: what are people saying about my brand? People were most interested in using this technology for anything data-related. While executing all these transactions, we also discovered some issues. The first problem was that you had to clearly specify the endpoint depending on the task you wanted to solve.

That is, we expected all agents or human agents or human operators behind them to know which endpoint to call. The second problem was execution. By the way, when I say " endpoints," I mean X.509-compliant endpoints—a X.509-compliant endpoints—a protocol that allows agents to purchase and pay for services based on usage.

This protocol is growing steadily, and now there are over 37,000 such points. But the problem is that many of them work very poorly. It often happens like this: you pay for something, the endpoint takes the money but returns an empty result. So, with these two issues and our experience, we went back to planning and started thinking: what can we build to optimize usage and solve these issues at the same time ? As this field develops extremely rapidly, these two problems are becoming increasingly significant.

So what if um, sorry. Yes. Going back to my previous analogy: what if a car had an intelligent engine that knew where to go? Then you wouldn't need a guide. What if you also don't have to worry about whether the restaurant you're going to accepts cash, cryptocurrency, or cards, whether it accepts dollars or any it accepts dollars or any other currency.

So how is all this processed by a car engine? What if you don't even need a wallet for your agent, as he can handle all these issues on his own. So, that's what we did, and we built a composite model. You may think of composite models as ordinary, but they are made up of different models, with faster models solving the simplest tasks simplest tasks and smarter ones solving the more and smarter ones solving the more complex ones. So, why create such a model?

As I said, it's like a car engine. So if she already knows what to do, you don't need to give her any more instructions. Plus instructions. Plus , because it's a model, it fits easily into your workflow, and you can connect it via MCP or API to via MCP or API to any client, whether it's Cursor or Claude. And so, how did we do it?

We trained it based on all the transactions it had as a starting point. So based on all of this, we know which endpoints are effective and for what, and because of that, the model knows that too. And then we take care of everything: from routing to execution and validation, hiding it from the user. So, routing or searching is the first part. Based on all these transactions, depending on the needs, we know which endpoint to use endpoint to use .

Then comes the execution. We remove all the complexity of working with cryptocurrency. So it doesn't matter the network or currency, everything just works in the background . And check. So, before providing the result to the user, we validate it.

If the confidence level is low or the result is empty, the user is not charged. It's already working and you can easily test it on frames.ag, as well as connect via the MCP API and use in your workflow. And with this single connection, you get access to all endpoints and beyond. You get all the routing, execution, and validation done for you.

We also worked on what we called a tool reliability system. Using all these transactions as a starting point, we see that the more the model is used, the more it learns. So, as transactions are successful or not, they get a rating, all endpoints are ranked, and the system gets smarter over time. This is all a bit abstract, so for easier understanding, here is a simple example. So, that's a hint.

The most important thing in it is just the first phrase. So, display the current option table for a specific ticker. The current options table is not something you can find through a web search. This is data that Google doesn't have data that Google doesn't have , and that's what Frames F1 is about. So, we're in the business of getting proprietary data, and this is a good example.

But, as I said before, it could be anything from brand awareness to lead enrichment. We have special tools for all this. So we took that tip and made it work with the Frames F1 and other cutting-edge models. Specifically, in this case—GPT-5.6 Soul and Stable 5.

And then we asked the models themselves to compare their results with Frames F1. And both models Frames F1. And both models essentially acknowledged that the Frames F1 result was better. This is simply because they do not have access to the data as they only have access to web search. And Frames F1 is designed for everything that goes beyond these limits.

Yes, I believe that in the next couple of years, every company will become available to agents, and the ones who will be more successful will not be those who offer their endpoints earlier or faster than others. These will be the ones that agents will choose by default. And if you sell data or have APIs you want to make available to agents, we can help you become the default choice. Thank you.