What 500,000 Settled Agent Payments Taught Us About Agentic Commerce

Transcript

[music] Hello everyone. So in 1900 there were about 3,000 cars in France. Two brothers who were selling tires. They decided to create a guide with relevant information for travelers. things like maps uh with all the roads, hotels, restaurants, and interesting destinations to get people to drive more.

Their goal was not to create a food ranking system. Their goal was to sell more tires. And they understood that what was missing was not infrastructure. What was missing was discovery and trust. I believe that in the agentic payments space we are in the 1900s but the solution is not to build a guide.

Let me explain. But first a bit of context. Like Cardc was saying, it's very nice to be here in Lisbon on the stage because we have a long history with Eve Global and we have been thinking about agentic payments for for a while now. First um it started in Elo Prague. The project was called MCPay and basically it was monetization solution for MCPs model context protocol already with aentic payments in mind.

Basically you have an API you make it MCP compatible and you monetize it using our SDK so agents can consume it and pay you as they use. This project became finalist in if global Prague and then with EF global support we decided to go allin. So in Delhi we did spotlight where we had a booth and presented the project to hackers and attendees of the event. And then in Buenosyus we did demo day where we also had a booth but this time we were pitching for investors and it worked out and we did the coliseum accelerator in early 2026

and then we also rebranded to frames. So now we're in early 2026. Agents are smarter. they can reason. We have skill.md files where basically you can teach an agent anything. But there was still not a straightforward solution to giving agent wallets, a wallet designed for agents. And this is what we built early this year. Very creatively, we named it agent wallet. And the cool thing about it was that it came with a cured registry of tools for these agents to consume. So

basically you give them a wallet and you give them tools for them to use these funds for. What sort of tools? anything from data providers, things like Twitter API, uh lead enrichment, um image video generation tools, trading, low latency trading tools, uh a huge variety of endpoints. The way it worked was it was prompt based. So it was just a single prompt that you could paste into any client like chat GPT or cloud and we took care of all the rest. So serverside wallets, passphrases, we abstracted all this complexity for the user

and it was very novel at the time. There was just nothing similar to it and it really took off quickly. So from then until now we have processed over 500,000 transactions. And what did all these numbers teach us? What did all these transactions show? This is what I want to talk to you about. So the first thing that became very clear was what the people were using this technology for. So what were the agents doing with this new access to wallets and tools and over 90% of all the transactions were data related

purchases. Like I said before, this can be for example Twitter API or lead enrichments or anything like brand awareness. So what are people saying about my brand? Just anything related data is what people were most interested in using this technology for. There were also some problems that we discovered along the way with all these transactions. First problem was you has you had to really specify which endpoint to use depending on the task that you want to solve. So basically we're expecting all the agents or all the human operators behind it to know which endpoint to call depending on

what they need. And second problem was execution. So right now um by the way when I say endpoints I mean X42 compatible endpoints which is just a protocol for enabling agents to buy and pay as they go and this protocol has been growing very steadily and there are over 37,000 endpoints there right now. But the problem is a lot of these endpoints are really bad. So a lot of times you pay for something, the endpoint takes your money but gives you an empty result.

So based on these two problems and on all the insight that we had, we sort of went back to the drawing board and started to think what can we design to first of all optimize for what people are using this for and also solve these two issues. And since the the space is growing so fast, these two problems just become bigger and bigger. So what if um sorry yeah so going back to the analogy that I made earlier what if the car has an

engine that is smart that natively knows where to go so you don't really need a guide what if it it also you don't have to care if like the restaurant you're driving to accepts cash or crypto or credit card if it accepts USD or any other currency. What if all of this is handled by the engine of the car? What if you don't even need a wallet for your agent because it natively can solve all of this thing? All of these things. So this is what we did and we built a composite model.

So you can think of composite models just like regular models but there are composition of different models where faster models solve the easiest part of the task and smarter models solve the more demanding part of the task. So why build a model? Like I said it's like the engine of the car. So if it natively knows what to do, you don't have to instruct it anymore. Also, since it's a model, it just easily fits into your workflow and you can connect it via MCP or API into any client, be it like cursor or be it

cloud. And yeah, so how did we do it? We fine-tuned it with all of these transactions that we had as a starting point. So based on all of this we know which endpoints are good and for what they are good and with that the model also knows that and then we take care of everything from routing to execution and to validation and we abstract all of that from the user. So routing or discovery is the first part based on all of these transactions depending on what he needs we know which endpoint to use.

Then comes execution. We abstract all the crypto complexity from it. So it doesn't matter the network, the the currency, it just works under the hood. And validation. So before we give the user the result, we validate it. If the confidence is low or if it's an empty result, the user is not charged. This is live right now and you can easily test it on frames.ag. AG and you can connect to it also via MCPR API and use it in your workflow. And with this single connection, you get access to all of the endpoints and more of that. You you get

all the routing, execution and validation done for you. We also worked on this what we called tool reliability system. So using all of these transactions as a starting point the more the model gets used the more it learns. So as transactions succeed or fail they get ranked all the endpoints get ranked and the system gets smarter as it goes. This is all a little bit abstract. So to make it a bit easier to understand here's a simple example. So this is a prompt. What really matters from it is just the first phrase. So put the live

options chain for a specific ticker. Life options chain is not something that you can find with web search. This is data that's not available on Google. And this is all that frames F1 is about. So we're all about fetching proprietary data and this is a good example for that. But like also I said earlier, it can be anything from brand awareness or lead enrichment. We have specific tools for all of this. So we took this prompt put and made it work with frames F1 and with other frontier models. So in this case GPT 5.6 Soul and Fable 5. And then we asked the models themselves to

compare their output to frames F1. And both models basically admitted that frames F1 output is superior. And this is simply because they don't have access to the data since all they have access to is the web web search. And Frames F1 is for everything that is beyond that. Yeah. So I believe that in the next couple of years every company will be accessible to agents and the companies that will have more success won't be the ones that

offer their end points earlier or faster than the other ones. It will be the ones that the agents choose by default. And if you sell data or if you have APIs that you want to make accessible for agents, we can help you become a default for that. Thank you. [applause] >> [music]