Agentic Commerce: AP2, x402, and Autonomous Transactions | Marc Vanlerberghe - Algorand Foundation

Transcript

And um I want to zoom out a little bit compared to what the previous speaker said and talk a little bit about Agentic Commerce in general and where X42 fits into the overall picture of Agentic Commerce, where we are as an industry today and where the gaps are and talk a little bit about how we look at those gaps and how we're trying to contribute to the overall XRO2 ecosystem by filling some of those gaps. And if you have some time at the end, I would love to actually do a demo where I show you how all of these different things kind of combine uh together. So, let's get started. This is an interesting chart, right? It shows Google Trends for Aentic Commerce. When we were here last year in February, there was nobody talking about a gentic commerce. There was it was not

a thing. Google search and Google Trends show pretty much zero people searching for it. And look at what happened over the last 12 months. It just exploded. It just exploded uh which is quite remarkable. So what is it at the very basic level? You know what if you could type into a Google search buy me these shoes at size 10.5 and be done with it. Right? At the basic level that's what the agent commerce what the gender commerce is about. But what people don't realize is what a fundamental different uh change and what a massive change this will provoke for pretty much everybody who's operating in commerce today. First of all, from a point of view of what is agentic commerce, the

big shift that is happening is moving from an AI that is assisted to an AI that is agentic. Right? We all use OpenAI. We all use plot. We ask it to do all kinds of stuff for us like, you know, uh summarize documents, doing research, helping us code. But these tools are not very good at at doing things, at buying goods, at buying services, at buying an airline ticket or doing any of those types of things, right? The change that is happening now is these tools are going to get better at actually doing stuff, at being a taking action, that have an agency. So what does that mean? I mean if you think about it from a commerce perspective it is a revolution right? If you think about what is happening to tra

traditional merchants, right? Merchants during COVID had a massive shift and try to invest and double down on e-commerce and optimizing their e-commerce infrastructures, right? All of the ecommerce infrastructure that these merchants have built is of no use in an agentic world. The reason why is because merchants have built their interfaces and their UIs for humans and agents don't care about about any of that, right? If you think about this as a merchant, first of all, you do search engine optimization, you do search engine marketing, you make sure your your store is lifted listed high on Google. that when people click on the link, they go to a tightly curated experience where the merchant is showing

you products and and positioning them and making them and making them as attractive as possible. They're doing personalization. They're showing you the products that most interesting to you. They're optimizing the checkout. They're optimizing loyalty after purchase. All of that stuff doesn't matter anymore in the agentic world because agents don't care about any of that anymore. So if you're a merchant, you should be really scared right now because you have to fundamentally rethink everything that you know about commerce. First of all, you have to think about building MCPs. So if you don't have an MCP and people start shopping through agentic AI systems, these AI systems won't even be able to talk to you as a merchant. So so you're not even in the consideration

set. So first thing you have to do is you build you have to build an MCP server so at least the agent can talk to you. But the problem is with the MTP server is that the agent will ask for details about what shoes do you have in your inventory, what do they cost, what colors, what size, etc. But then it will compare your offer against the offer of other merchants and it will pick the cheapest one. So as a as a merchant, you're now not doing a good job at like creating maximum value, right? Because the agent will go for the cheapest one. So as a merchant, not only do you need an MCP server, you have to build a whole agentic environment yourself so that when an agent talks to you, you have an agent that can talk back and that can start negotiating and can start doing

the things that are important to you. So it's a fundamentally different way of thinking about commerce as we go forward. So that's only one step. The second step is agenting commerce will also enable completely new markets. Imagine you want to buy, you want to order a trip, you want to organize a trip to Europe, right? You'll have a network of agents that will help you. Some some agent will be focusing on airline tickets, the other one on booking hotels in Paris, some other agent may be specialized in rental cars or in booking a hotel in Madrid. And so all of these networks and all these agents will become an ecosystem that will help you achieve achieve a goal.

That's a new kind of opportunity. The second thing that will happen is you and we see this with global. You have billions of agents that are all operating that will create marketplaces. They won't they won't just create social networks like what we've seen over the last couple of weeks. They will create marketplaces and in these marketplaces they will exchange products and services and goods and they'll have to pay each other which is not an entirely new opportunity as well. And then thirdly, you have the emergence of microp payments obviously and the previous speaker spoke about that a lot and this is where crypto can add a significant amount of value because microp payments are just not feasible over traditional card networks.

Okay. So how does crypto fit in? As I mentioned the previous speaker did a good job at that so I won't dwell on this but of course crypto becomes a payment rail for those microtransactions which is something that traditional agents are not traditional card networks are not really good at. Uh the fact that we have program bankability and smart contracts allow you to do fun stuff that that that again is very very hard to do in traditional finance. We talked about transparency and trust and traceability. Okay, this is an interesting one. Emerging standards. The oldest standards on this page is 14 months old and it's MCP. All the other standards are just months old, right? A2A was launched in April of last year. ACP X42 was launched

in May. AP2 was launched in September. UCP was launched last month. So all of these protocols, it shows you how new, how new and how emerging this marketplace is. And I want to basically spend a couple of minutes to kind of position all of these things together in the context of a commerce. How do these kind of pieces fit together? And this is an architectural overview. So imagine you have two agents. You have a shopping agent, you have a merchant a merchant agent. You want these agents to be able to talk to each other, right? One agent may be Gemini. the other agent might be open AI. How do they talk to each other?

This is where the A2A protocol comes in. A2A basically enables two agents to talk to each other no matter what the underlying LLM is. An agent needs to have access to wallets. How does an agent get access to wallets? That's where MCP comes in. MCB basically abstracts the blockchain away and gives agents a skill to interact with the blockchain. So the agent can basically have access to a wallet and do things on chain. That's MCP. Then you have AP2. What is AP2? AP2 is a layer is an extension on top of A2A and it basically creates the idea of uh intent mandates of mandates. How do you know as a as a user that the agent that you've instructed to buy something on

your behalf that the agent is doing the right thing? What's the audit trail? How do you how do you match the intent of the user with what the agent has actually done? That's what AP2 solves. They create mandates. So as a user, I will give the agent a mandate. will cryptographically sign the mandate. The agent executes on the mandate and afterwards you can compare the results. You can see whether the mandate matches what the agents has effectively done. So it basically creates an audit trail for commerce. So that's what AP2 does. Then we have ERC 804. This is the registry uh for agents. How do agents discover each other? Like it's nice to say you have a shopping agent talk to a merchant agent, but how do they found

each other? Right? This is where ERC 80004 comes in. It's a discovery mechanism. So agents can be registered. Agents can have identities. Agents can build a reputation. And it's a way to basically organize and discovery of agents. And then X42 finally is a method that is used to execute payment transactions on chain. So there's a lot of open questions and we'll talk a little more about that around security and identity and frameworks. uh but this is the basic landscape of how agent commerce works and where X42 and crypto fits into the into the overall landscape.

Okay, X42 native payments the previous speaker spoke spoke about that as well. So so I'll be very quick on this. This X42 protocol was launched by Coinbase uh last year and the idea is there's basically three parties in X42 system. You have the um uh the the the buyer, you have the seller, and you have the facilitator. The buyer and the seller don't need to know much about the blockchain. They don't really interact with blockchain. They communicate over HTTP and they exchange basically payment instructions. The facilitator is the third party that will then take those payments and then settle them on the blockchain and make sure that everything is happening correctly. So that's how the system basically is is organized.

So there's a bunch of issues with this. U one is there's the goldilux problem, right? If you have a buyer and a seller, for this buyer and the seller to be able to interact with each other and to make payments, they have to be on the same chain and they have to support the same currency. So So if you have one, you know, uh agent that's on base and the other one is in Salana and one is USDC and the other one is Tedar, it won't work. It won't work. So you have this goldilocks problem. How do you make sure that that that that's a big problem that the industry basically has to solve.

Delivery versus payment. The previous speaker spoke about that as well. There is no chargebacks. There is no mechanism to ensure that what was paid for was actually delivered. This is not part of the X42 protocol as of as of yet. There's economic sustainability. How do these facilitators charge? Um how how transparent is the charging mechanism of these facilitators? How can you do fee delegation so that maybe somebody else is charging for the transaction on behalf of the facilitator? Um, payment flow flexibility is another big issue.

Currently, the payment flows are very simple. It's like person A sells person to person B and there's one transaction that there's a payment transaction. But payments can tend to be much more complicated than that. And when you think about this, you could build very very complex payment instructions on how a payment gets gets gets consumed. And currently X42 doesn't allow that to happen. discovery and trust as I mentioned ERC 80004 is discovery mechanism for agents but it's kind of like the Yahoo of agent search right now right like remember the old people in the room like myself remember like how Yahoo used to work on the internet it was it was basically a directory service right there was no Google you didn't search you went to the Yahoo directory

and the directory gave you the you know the phone book of websites where you could actually and the links to get there right that's what ERC804 basically is it's a phone book of agents that shows you what agents are available. But that's kind of the agentic Yahoo. That's kind of the stage we're in. We need to move to an agentic Google where where agents can be discovered in different ways. Security. How do you handle security, right? How do you make sure that the agent that has access access to to a blockchain is doing exactly what it needs to do and is not going wild and is not overspending. What what kind of mechanisms are in place to do that?

Nothing in X42 deals deals with this. And then finally, probably the biggest one is user experience. It's very easy for an endpoint to create an Xro2 enabled website, but what is the other side like? How does a user interact with that endpoint? How does the user and and if you want this to be be available to more than a cryptonative audience? How does this scale? How do you create a userbased infrastructure that users can easily pay for those microservices in a way that they don't need to know about blockchains or need to know about wallets and private keys and any of that stuff. This is a really big big issue as well. So let me talk quickly about the algrand approach to xro2. So um we we

are very excited about this because I think our chain has some characteristics that are really well designed for those microp payments. Our fees are very low. We support instant finality. Our chain never forks. Every transaction that gets included into a block is automatically confirmed. It will never be reversed ever. And so block confirmation is less than 3 seconds. So it's a chain that's really well designed for for those types of microp payments. And so the biggest and most important feature in this context for X42 is that our X42 spec that was merged recently by Coinbase has a concept of native atomic grouping. So now payment transactions do not have to be single transactions.

They can be group transactions that are grouped together. And the and the interesting thing about that is you now establish really complex payment instructions and ensure that either all of these instructions go through at the same time or they all fail at the same time. And this is a native feature of Algarand. It's called atomic grouping. And there's all kinds of ways that you can deal with this. Basically, it guarantees it bundles all these transactions and it executes them within the same block guaranteed and they either all go through or they all fail. And so you can imagine in the context of like the Goldilocks problem I mentioned before, right? If the seller and the buyer have different currencies, maybe the payment instruction in the middle does an automatic swap from the seller currency to the buyer from the buyer

currency to the seller currency. You could do that with an atomic group or one the seller is on one chain, the buyers on another chain. Well, maybe you can bridge right the payment across chains as part of an atomic group. So, it opens up a lot of possibilities to kind of deal with this goldilocks problem that that that we currently have. The second one is security um and account abstraction. So, we've uh launched a feature called ARC 58 that tries to deal with security onchain security. And this is all about making sure that the agents have constrained autonomy. They can do what they need to do but they do it in a constrained manner. There's two core elements to that. One is the type of transactions that an agent can do is constrained by a plug-in system. And a plug-in system is

basically a system of smart contracts. Each plugin is a smart contract. And the and the agent can only interact with these plugins. And so therefore, it's constrained in the type of activity and the type of actions that it can undertake. The second thing we do is agents never have access to funds ever. the when the agent wants to make a payment, the agent basically gets temporary access to execute one payment instruction and then the access is is is reverted. And so we do this again using this idea of atomic grouping. And so it's called flash rekeying. So we have this feature called rekeying where you can temporarily hand over control over an account to a different account. And so what you do you at the beginning of an atomic group you rekey control over

the spending account to the agent. The agent executes his transaction constrained by the plug-in and at the end of the atomic group you take away that access to that spending account and you bundle it into one atomic group. So at the end of that transaction before the transaction the agent has no control of the spending account. At the end of of spending accounts. Everything is handled within within that flash atomic group. So let me actually show you a quick demo um to end this presentation. And this is a demo that we've built um together with Google to to demonstrate how all of these different pieces come together. Uh how do you combine A2A, MCP, AP2

um um and X42? And so let me play this and let me show you what happens. So on the left you have a shopping agent. On the right you have a merchant agent and they're basically going to interact to buy shoes. The first thing they do is they connect to an MCP server and the MCP server makes you log in. So there therefore the agent now has access to a wallet. So the agent now has access to agon blockchain, can look up wallet infrastructure and everything. So we'll check that everything works. We say hello to the shopping agent. We say hello to the merchant agent. And so everything works fine. So then we basically add a shoe to a cart. Then we say let's go buy this shoe in size uh 10.5.

There we go. So now what's happening is the merchant, the shopping agent is figuring out who do I go to, which store do I go to. So it basically is getting a merchant list. For this demo, there's only one u there's only one that's available. Um and then it's getting the merchant card. So now we're in A2A lands. As part of A2A, every agent has a merchant card. So you can look up the merchant card and the merchant card has information about what the agent does and uh what its capabilities are and how to most importantly how you contact the agent.

So it's getting the agent card. It's checking the wallet to make sure it has access to the wallet. And then it basically generates and here we go. It generates an intent. So it creates an intent. Now we're in AP2 land, right? So the agent is basically saying, "Okay, this is the type of shoe you want to buy." I'm creating the user intent. I'm cryptographically signing the intent and then I'm actually going to package that intent which is AP2 inside an A2A message and send that A2A message to the merchant agent. And you see on the right that that message just arrived on the merchant side.

Okay. So now the merchant agent is searching for his catalog and it finds it finds the tool. It finds the um uh the shoe and it's now preparing the payment instructions to the shopping agents. So you can see it basically looks up the uh facilitator the X2 agent card is basically the facilitator. It looks up the facilitator to get information about how that works and it is constructing a message an X42 payment required message that provides all the details about what payment is required like you know the currency the chain and all the details that go along with the payment packages up in an A2A message and sends it back to the shopping agent.

Okay, shopping agent gets it is checking shipping addresses and then is generating a payment transaction. So now it's taking the payment instructions from the merchant agents and it's creating a payment uh a payment transaction. It is signing the payment transaction with the private keys of the user and now it's creating this pay payment payload. So it's going to send back an X42 message to the merchant saying okay I'm ready to pay. Here's my X42 payment. It's basically a payment transaction that's fully signed that's transmitted not on chain but that's transmitted back to the merchant agent. So now the merchant agent is receiving this. There we go. And the merchant agent now will contact the

facilitator. And so um it needs to send the X42 operation to the facilitator. It's getting the facilitator card. There we go. And now it's basically asking the facilitator to verify this transaction. Does this transaction look good? Does it have the right instructions? Does it have the right amount and everything? After it does that, it asks the facilitator to settle and this is where the facilitator will interact with the blockchain. This is where the facilitator will actually submit that transaction to the blockchain. There we go. And the payment was just uh completed. So the facilitator comes back to the merchant agents and says, "Great,

the payment went through." The merchant agents contacting the shopper agent and saying, "Great, you've just bought those shoes." and the shopper agent is then producing a receipt that the user can then look up and uh validate using any kind of blockchain explorer that the transaction actually went through. So that's a lot combined into one little demo. Uh so you can see how all of these things started fitting together. You see the block explorer here. This is where the actual transaction then shows up. Um, so this gives you how basically this this tells you how the entire system kind of works and how all the different components kind of combine together to create an end to end agentic commerce

flow. And that's all I have for you today. Uh if you want to know more about how we uh uh build extra two products in Augant, check out this QR code and uh and happy to talk to anyone uh after this.