Agents That Can Pay: Circle & Mastra

0:22 Hey, hey, hey. >> >> Hello everyone. Good morning, good afternoon, good evening wherever you're calling in from. message in the chat where you're joining from today. We're super excited today. We have a really great session. I per usual am calling in from San Diego, an abnormally humid San Diego today. We have a lot of tropical storms around, so a little bit more sweaty than I'm used to, but still still a beautiful place. And then Anthony, where where are you calling in from?

1:43 >> Calling in from LA. >> LA. We have a big SoCal contingency today. >> Yeah. And we have Kainer from Berlin. Awesome. Good to see you. I hope everyone that celebrated Labor Day this past weekend had a good long weekend. if not, let let us know in the chat what you got up to this weekend. Maybe if you built anything cool on the side. I think today always try to ask a a question about MRA to try to get the chat going. I'd love to hear the first agent you've ever built with MRA in the chat.

2:20 or maybe you're just here learning and you haven't built anything yet and your first agent can be an agent with a wallet, which would be awesome. >> Let's see. And we'll get started in just a few minutes. Just going to wait for some people to roll in. >> >> Nice. Lots of lots of emojis coming in.

2:57 I didn't even know that was a thing. I don't think I've seen that past workshops. >> Yeah, they they show up on your end, too. >> Yeah. >> Love the support. And as you're joining, we're just going to get started maybe like a minute or two now. I know a lot of people are are coming off the holiday weekend as well, so I want to give them a little bit of time to get to their computers.

3:26 Maybe a later start this morning. Aaron, first agent analyzing calls transcripts. I remember that agent well. Very helpful one. Yeah, Aaron is also facing from from Hawaii. He's facing a lot of the tropical storms as well. Probably a lot a lot worse than us. For for me and Anthony, it kind of just manifests as a little bit of rain, but I know Hawaii has dealt with a lot more intense storms.

4:15 The plants are loving it. Yeah, that's the big thing. Every time it's kind of like a joke, but every time in San Diego or LA, I feel like it rains, everyone's response is just like, "Yeah, the plants needed it. The grass needed it because it's always so dry out here." >> but yeah, I'm sure the plants are are having the time of their life. We'll just give it one more minute here and then we can get straight into it.

4:45 I know we have a lot a lot to go through today, so I want to make sure there's a lot of good time. All right, I think we can get started here. Well, hello everyone. Hope you're all having a great day wherever you're calling in from. I'm Brandon. I host this workshop every week with different partners and customers that are using Maestra and we dig into different use cases. Today we have a session that I am super super excited for. I know I've told Anthony this many times, but this one I've really been looking forward to.

5:23 Anthony is from Circle and he's going to be talking about agents that can pay, giving agents wallets so they can complete different transactions. obvious a more and more popular use case for different agents and is something that I've heard a ton in different calls with companies I've been working with and figuring out how to do this safely, securely, have a lot of auditing around it, all of that stuff. so with that in mind, just want to remind everyone, you know, feel free to use the chat as much as you want. Feel free to talk with other people. You can put any questions in there. I'll try to get to them either throughout the session as they sort of become relevant or at the end we always save maybe five to 10 to 15 depending on how deep you guys want to get minutes for questions at the end where we can all chat through things. So with that in mind I'll pass over to the real star of the show Anthony and you can get started with talking about agents that can pay.

6:19 >> All right thanks Brennan. So hey everyone my name is Anthony Colani. I'm the director of ecosystem growth at Circle. I work all with developers technology partners to grow adoption of circle's developer platform and tools. I'll start off a little a bit talking about circle. circle you know circle we're building the foundation of the internet financial system. at its simplest level we're known for USDC the most widely used regulated stable coin in the world but the broader vision is much bigger than issuing USDC. you know we're building a full stack platform of for internet native finance across digital assets developer infrastructure and apps. so as you know software has always been great at moving information but moving money is a is a much different story.

7:08 the world's financial infrastructure has still has a lot of friction and and our view is that money should move with the speed and programmability and openness of the internet but also preserve trust and reliability and compliance. So, you know, we built USDC, we built, you know, our circle developer platform and ARC, our layer 1, upcoming layer 1 EVM compatible blockchain to make it a lot easier for developers to build serious financial apps that can serve customers globally. So, over the past few years, we, you know, we've seen how agents are becoming more and more capable of doing real work. they can research, they can perform really complex reasoning and workflows, but the moment that they need to pay for something, they often get stuck. most payment systems were built around humans with accounts, you know, password, credit cards, API keys, things like that. this creates a lot of friction for autonomous software and makes really small like per request payments really difficult to offer economically.

8:08 So for circle agent stack, you know, it was built to basically close that gap. it gives agents infrastructure to discover services, hold USDC, make payments to services that an agent might need, you know, to complete a task, but also giving humans that run the agents like clear polic control over like the policies and and kind of approval controls for those agents. So today I'll kind of show you how that works with Mastra. we'll follow an agent basically with scenarios like it will have 20 minutes to prepare you know a partner brief the agent will manage his own wallets it will find services on agent circle agent marketplace evaluate their costs ask for approval pay in USDC using x42 nano payments which we'll go over and this is all done using our circle agent stack and master's SDK so agents aren't just using the internet anymore. They're becoming customers of it. you know, but machines participating in commerce isn't entirely new, right? We have ad platforms that bought and sold ad space. We've had trading algorithms that make decisions and move value across the internet, but those systems operate more so in a closed kind of purpose-built environment.

9:29 some of like the participants and rules and like the relationships, commercial relationships are established in advance. But general purpose agents are very different. you can give an agent a broad objective and it can decide how to accomplish it. You know, halfway through that work, it may discover that it needs a capability that it's developer never anticipated. Maybe it needs some type of specialized company data lookup or it may need some compute or it may need like some very specialized analysis that's behind like a payw wall or some type of service you have to pay for. you know, an agent doesn't necessarily like in that moment want to sign up for a subscription or establish some type of long-term, you know, vendor relationship. It just needs the useful result like right at that moment. on and on the other side, a developer may actually have that exact capability. They may want to offer it to agents and publish it as a service, attach a price to it, and then get paid whenever an agent uses it. So, this is kind of like the basic shape of the agentic economy. Capabilities that can be published, discovered, priced, and purchased on demand even when a buyer and seller had no formal relationship before the the request began.

10:43 But the internet was missing a a native way to actually make this exchange happen. Back in 1996, HTTP reserved a status code called 402 for payment required, but it never actually defined how a payment should actually work. So, it kind of sat you know, unused largely for almost three three three decades. Then in May of last year, Coinbase introduced X42, which is a open standard which was established along with collaborators, you know, including us, Circle. and the idea was to make that missing payment layer actually useful. So instead of redirecting an agent to a checkout page, a service could return a 402 response containing machine readable payment instructions. the agent could evaluate that price, authorize a payment, retry the request, and then receive the resource after paying. So this is all within the normal kind of internet protocol HTTP interaction. And so by April of this year, you know, we've been seeing early signs of a eco an actual ecosystem forming. so now we have, you know, approximately 70,000 active paying agents, could be more today actually. the number keeps rising, you know, more than 165 million XO2 transactions and with an average payment of about 25 cents.

12:09 so you know, the number of paying agents really shows that there is a real ecosystem of buyers. The transaction volume really shows that like there's repeated usage, not just like these isolated demos. And then, you know, we're seeing like 25 cent average tells us like what kind of economy this is becoming. It's more so like high frequency, low value purchases that traditional payment rails were never designed to support really efficiently. like you, you know, 25 cents doesn't economically make sense for for a for a credit card transaction, but it can be very valuable when an agent is purchasing precisely the data or compute that it needs to to finish a task. This market is still early. you know not every transaction represents like a fully autonomous agent or like a mature use commercial use case but the underlying structure is emerging. We have buyer sellers pricing discovery and value moving between different software systems and XO2 gives those systems a way to have the actual payment conversation. But the next question is is like what form of money works best inside of that conversation. So we move on to USDC.

13:21 so it's not enough just for money to be digital. It has to actually match the economics of agent to commerce. So imagine an agent paying 4,000 of a of a dollar for every query it needs from a service. It if it makes a thousand of those requests, that's $4 in total. But the economics only work if each individual payment can happen efficiently. So that creates a a really particularly like a very particular set of requirements. payments needs to be fast enough to keep the workflow going.

13:52 You don't want to be sitting there for minutes waiting for agent to finish something. it needs to be inexpensive enough that the that the actual like transaction fee doesn't exceed the value of the service. So, if you're buying a service for like 1 cent or a fraction of a cent that the actual fee to do that can't be over that because it doesn't make economic sense. they need to be able to work across borders and platforms and there also can't be any kind of like artificial commercial minimum. The price should really reflect the value of one query, one inference or one task. USDC fits those requirements really really well.

14:27 it gives buyers and sellers a dollar denominated unit of account. an agent can re reason about its budget in familiar terms, US dollars, you know, while a developer can price a service without introducing the volatility of another type of digital asset like Ethereum, which goes up and down in value. You know, we're we're pricing things in like US dollars. So, USDC can be held and transferred programmatically across different blockchain networks. this means the payment can become part of the software, the actual software workflow, not a separate checkout experience that requires a person or card or some type of banking relationship.

15:07 And we can already see that developers are making that choice. Like today, 99% of X42 transactions are settled in USDC. you know, that doesn't mean that the market has reached its final form, but you know, you know, agentic commerce is still developing quickly, but it's a strong signal that, you know, when developers need a settlement asset for agentic payments, they're choosing USDC. And it makes sense. You know, USDC gives agents money that behaves more like software. It's programmable. It's composable. It's available online. It's predictable and valuable in value.

15:41 you know but and but giving an agent access to money is is really only one one part of the problem. It still needs a wallet, still needs like spending controls, service discovery, things like that. And this is where agent stack comes into play. So agent marketplace supports discovery. It gives agents which is the buyer side a structured way to find machine callable services while giving developers the seller side a place to make their APIs and capabilities discoverable. Agent wallets provide a controlled you know control access to funds. Agent can hold and use USDC but only within the policy established by the developer you know the person or organization responsible for that agent.

16:26 nanop payments supports very small highfrequency payments. it makes it practical to to charge for individual query or inference or task and settle those payments efficiently at scale. And then connecting the workflow is is circle CLI gives agents a a real deterministic kind of command service surface for actions such as like checking your wallet balance, discovering service or initiating a payment. And then we have circle skills which provides the instructions to the agent on how to actually work use circle CLI. We'll dive into that a little bit in more detail in a second.

17:04 But to together these components support both sides of the market. The buyer and seller side. A buyer can discover services, evaluate it, pay for only what it uses. the seller can actually publish a capability, set a price and receive USDC without, you know, the the buyer having to create an account or get an API key, things like that. so overall s agent stack provides the finan financial capabilities and then mastra turns those capabilities into a working agent.

17:37 So let's go here. so this is best so the integration is easiest to understand kind of two layers connected by like a control interface. it begins with the user's objective in our demo that objective is to prepare a useful partner brief before a meeting. master orchestrates the work around that goal. It gives the agent a workspace and terminal. it manages its reasoning and tool calls, manage maintains the memory, you know, records the run so we can actually see how each decision is made. So it provides tracing. you know when the agent actually needs financial capability, it turns to circle skills. the skills don't actually you know hold the funds or execute the payments. They give the agent instructions on on how it can use the tools to complete the task. So you know like we said checking the wallet balance, searching marketplace, things like that. Those instructions lead the agent to Circle CLI. The CLI provides a consistent command interface into Circle agent stack. So instead of asking the model to construct, you know, its own API calls, its own signing logic and building all these things on its own, the agent, you know, works through a set of defined commands through Circle CLI.

18:58 and behind those commands are the f financial capabilities we just we just covered. Agent wallets can control agent wallets. It can do things in marketplace. It can also do payments through nano payments. so just kind of reinforcing this the division of responsibility is clear. Master orchestrates the task while agent stack provides the financial infrastructure for the agent. >> >> Yeah, Anthony, I'd love to jump in here for a moment because I think this is a really great slide to address one question I've gotten, which revolves around how do you sort of make outputs determin deterministic enough for money?

19:35 Meaning, how do you develop an agent that is naturally non-deterministic and start to trust it with money? And I think this is a really good slide to kind of show how master and circle can work together here from the MRA end. If you have it as the orchestration layer, this is where you can add in a lot of the logic that goes around, you know, maybe having a workflow where there's human in the loop before you're using circles capabilities, having the auditability as well so that you can track different things that have happened. And I don't know if there's anything you'd like touch on on like the circle end for that kind of stuff, but I think one of the biggest concern as people get into payments is, you know, I I've seen agents kind of do different tasks that I didn't intend for them to do. How do I make sure that once I'm putting money in the mix, there aren't those same risks?

20:21 >> Yeah. Yeah. So, like, you know, when it comes to agents and LLMs, obviously, we know that models can come up with like really creative answers and and they can reason and and make, you know, decisions on their own. However, when it comes to moving like money and value on chain, we don't want like agents just randomly doing things and going ro and having the possibility of going rogue or even messing up. So there's multiple layers of like determinism that we've established on the circle side to make sure those those those things don't happen. So with the CLI it actually creates like I said it creates a a deterministic like set of commands that an agent can use. But in addition to that you know there's several layers of capabilities you can kind of program in there. You can tell agent to be careful with money that is not fully foolproof.

21:14 then you can, you know, programmatically maybe make some like guard rails around that. You know, that works to a certain degree. But then at the actual wallet layer, that's what we've done on the circle side. At the wallet layer, this allows for a human to actually create policies that the wallet enforces. So, say an agent makes a mistake and wants to spend a dollar or $100 in USDC rather than a dollar. If you set a policy on that wallet that says you can't like that this wallet can't transact over five USDC in a day that when that agent tries to use circle CLI to make that transaction that wallet will like basically like make sure that transaction doesn't happen deterministically right so that's those are the capabilities or the things that we've built to like make sure that agents have the freedom to spend money and and make their decisions but allow a human to actually like put policies and put controls around that so a agent doesn't go rogue and do something that the that the developer didn't intend for it, you know, to happen outside of its parameters and it it policies. So, I don't know if that answers the question, but >> yeah. Yeah, 100%. I I think as you said, there's sort of all these different guardrails you can play around with and yeah, feel free to keep going. And I just I saw this slide. I'm like this is I feel like a great one to sort of bring a little bit of understanding around something that is you know naturally from an enduser perspective a bit of like a black box but I think the way you've laid it out here and also chatted through circle stuff really shows a lot of you know there is a lot of tangible deterministic guardrails that you can put around it to make this a really safe reliable and trackable experience.

22:58 >> Yeah and I'll show that in a demo as well. Just an example of that. but yeah, that was pretty much what I was going to go over the rest of the slide is is you know the the different layers of controls that can be put to ensure that you know agents you know have you know safety around like you know spending actual real money you know. So so yeah move on to the CLI. So you know as we said you know agent frameworks are are really good at generating you know reasoning and creative answers and things like that but financial operations should definitely be less creative they should be deterministic. you know agents also shouldn't have to invent different you know integrations every time it needs to to do some type of wallet you know operation creating a wallet discovering service or making a payment it it needs a consistent vocabulary. So this kind of this slide shows a little bit about you know circle CLI what it provides. It provides a single command line interface as you can see here on the slide. you know there's a few command examples here. We we're creating a wallet. there's a command here to search the marketplace. It returns the services and the prices and things like that. And then this last command here is like for you know paying for a service. we're passing, you know, the the endpoint of a service that needs that requires a payment to to the the CLI and the CLI is actually going through all the steps to make sure that payment actually happens.

24:33 you know, the the interface I mean extends beyond these examples here. it can work you can actually use the CLI for altering policies and things. there's some additional like security around that. So agents can't change their own policy, but a human can use the CLI to change policies. there's other u products that can be used as well like crosschain transfers using some of our you know our onchain capabilities or products that that we have at Circle. But you know now that we've seen the interface, let's kind of look at the kind of first capability behind it and that's agent wallets.

25:10 So for an agent wallet, the most important feature isn't isn't how much it it can do. It's just precise. It's more so precisely how we can control what it's allowed to do. And we kind of went over this already, but you know, Agent Wallace lets agents hold and transact USDC without a human having to cosign every single action. But really, that freedom only exists inside the policies, you know, created by the wallet owner. These policies can include, you know, a maximum amount per transaction, rolling daily, weekly, and monthly spending limits, allow list and block list for recipients and the different kind of contracts and and apps that it can interact with. you know, the wallet can still support, you know, useful onchain activity, payments, transfers, but really only within the boundaries that the owner has has defined.

26:00 there's some, you know, some examples here of values of what a policy might look like. here you can see, you know, oh actually for today's demo, you know, I set a policy on the wallet. So I actually set a policy that's a little bit more tighter than this. I set one that no individual transaction can exceed 5 cents and the wallet can't spend more than 15 cents during the active rolling daily window. I configured this policy before the the presentation. and the act the agent actually can't access or can't increase this limit or remove the policy or or even approve a policy change on its own.

26:40 So, it really needs a human to actually do it and we can do that through CLI in a different way. but this distinction between a policy and like other guards is is really important to understand. You know, pro a prompt can provide guidance. And you can tell it to be careful with money, but the wallet policy provides the actual enforcement. Let's see here. Let's go to the next slide. So, let's go to market agent marketplace. So, you know, when humans buy software, we often begin with a company or product we already know.

27:12 Agents are more likely to begin with a capability. So, they might search for like, you know, research a company or generate an image or some type of capability. The agent cares less about like the company and logo, but it cares more about like what whether a service can help it complete its next step in the task. Agent marketplace makes those capabilities discoverable in a structured machine readable format for agents. On the buyer side, an agent can understand what a service does, what inputs it expects for that API and what it returns. And then also like you know what the payment meth you know what payment method it supports you know and how much each request costs. On the seller side marketplace turns any API endpoint into a storefront for agents.

27:57 So you can describe what the endpoint does. Publishes like I said publishes inputs and output schema set a price and make it discoverable to agents. A new buyer doesn't need to create an account or get an API key or anything like that. It can discover the service and pay on demand for one single call. this creates a a brand new option between you know subscribing for software or building it yourself. It creates a whole new like business model for you know for API you know services you know and an an a an agent can acquire a capability for one task and use it and then just move on. in our demo, the agent will use marketplace to look up, you know, for services for information that service sources of information that are are complimentary to each other and not spend money on buying several different versions of the same answer.

28:50 So, we're going to generate a brief. We're going to look through all these services and then make decisions on which sources that we want to use. once it finds the right service, the next challenge is paying a very small amount for the usage efficiently. And that brings us to circle nano payments. So at this point the agent has found a service. It understands what it you know does how much it costs. The final question is whether the a payment of 4,000 of a dollar can happen without the transaction cost being more than the purchase itself. And that's where nano payments solves this problem. x42 is the payment conversation between the buyer and seller. USDC is the asset being exchanged and circle down a payment, you know, it helps make frequent low value payments economically practical. And then the CLI is how a agent accesses that workflow. The agent first requests the paid service. Instead of returning the resource, the API or service responds with a 402 payment required and includes a price quote.

29:54 Before spending anything, the agent can inspect the requirements, check that the purchase fits the policy of his wallet. And once the payment is approved, the flow kind of has, you know, three main stages. First, the the wallet signs a a authorization for the requested amount. This signed authorization is for a very specific amount and it can that that can be sent to the service from the agent wallet. So, it's almost like signing a physical check saying, you know, you have authorization to withdraw this amount from my account. next, circle nano payments verifies the signature and confirms that the agent has sufficient funds and then deducts the authorized amount. Once the payment is verified, the merchant can return actual requested data or service immediately. So, you know, basically request the service, you get back instructions on how to pay, you authorize the payment, and then when it's a when it's verified, then the resource and data that you requested gets returned.

30:57 so, you know, the individual per purchase is still priced and authorized per request. the but the onchain settlement actually happens later with nano payments in batches. And so this is important to understand because it changes the economics. Instead of turning every fraction of a cent payment into a separate onchain transaction, so we don't settle the payment every single time because there's a fee involved every single time we settle the payment. Nano Payments combines thousands of of payment authorizations into one settlement. So at the end of the day, you just settle once and you pay one fee. so the onchain cost is spread across the entire batch. so the agent actually doesn't pay a fee for each individual purchase. The fee is you know paid in bash at the end of the day.

31:46 That makes it practical this makes it practical to to charge by API to by per API call and you know the model response in the in the second of compute. So like these units of of compute like it it makes it really practical to charge per per call. with USDC payments as small as 1 millionth of a dollar. In master studio we can observe the entire decision process. the service request the stat the estimated pri price the approval every single step in that chain. We can actually observe that in master studio.

32:24 So now that we've covered all the components of agent stack, discovery, policy control wallet, agent native interface, and efficient per request payments, we can put them all together now and watch the agent use them. So let's give an agent a job that someone might actually delegate. Let's let's start with this demo. so imagine I'll go over the scenario. Imagine an unexpected partner meeting appears on on my calendar and it starts in about 20 minutes. I don't want another generic company profile. I want a concise brief that helps me have a better conversation with this com company. I'll give a agent a prompt and a task and that is pretty straightforward.

33:09 you know, research a company using up-to-date information, you know, identify where there may be a product or ecosystem fix. I want I want really useful information. And so, you know, this agent can use free sources whenever it needs to. It can search the internet. But if a paid marketplace service could improve the actual brief, it can discover that service, you know, price it out and propose a purchase. let's go to the prompt here. So, this is the prompt that I'm going to put into the agent. and maybe now we can switch over to the demo. So, I'll change my screen here.

33:49 Yeah. And real quick, while you're switching over to that, we can just do a catch a quick question. but someone has a question on the marketplace. So, is there any quality control or ratings or anything like that on marketplace services? Why would an agent feel like it can trust these services that it finds since there's no human that's curating it? >> Yeah. So, that's a good question actually. that's that's a big problem that's being solved now is is making sure services actually do what they say they're going to do because obviously payments involved in that. yes, we've started putting some real tools in there to like basically test these services out and and actually we've created a scoring algorithm that gives buyers a a a clear like basically a secure a clear score to see how reliable these services are and allow the agent to you know assess whether it wants to use the service or not. so yeah we we we're a lot of services are working on that. I know for marketplace we've been working on u making that making those scores very very clear for agents to to kind of decide like okay you know the service how how much has the service been up does it does it actually return quality results things like that. So yeah we're we're we're definitely putting those tools out.

35:16 All right. so, just a little bit about this demo here. So, we actually have created a an agent running on master server with a with a workspace. this is deployed in the cloud. this actual application you see here is just a it's a custom an X.js app that connects to that agent in the cloud. we just wanted to create, you know, a unique just a just a a custom UI just to show this demo so you can kind of see each step and you know, what the agent is doing. I set up a Circle Wallet prior to this demo. I actually funded it with like real USDC.

35:56 So, I own ramped, you know, USD to USDC. I I've already signed into the wallet. I supply USDC. I set up a policy which I described earlier. But what I'll do is let me create a new chat here and just kind of show you you know what's going on here. So let's see let me ask it you know my wallet comments and here it's using the the skills to you know check for the wallet you know check in to see if I'm signed into the wallet.

36:34 It's listing my wallets. It's going to check the balance. So, we'll get the balance here in a second. And then next, what I'll probably do here is so as you can see here, I have 5.223 USDC and a couple other tokens in there. what's the policy? I just want to show you the policy on my wallet. And I just want to show you like this is checking this is using Circle C allow to check the policy on the wallet to see, you know, what my spending limits are. Just want to show you like what we've set here. So per transaction I can spend only 5 cents per transaction. I can spend a maximum of 15 cents daily. And then there's a couple other weekly and and monthly limits here. I'll show you how how this policy works and enforces these limits.

37:21 So send let's just say 100 USDC. This is real money on on actual mainet too. And I'll just like get my wallet address here. So, it's not going to allow me to do this. it's going to try to, but it's not going to allow the wallet policy won't allow it to do it. So see here it's reviewing the spending limits and I it says I can't use send one USDC or 100 USDC because you know my limit is 5 cents per you know transaction. But if I do send let's send a fraction of a cent should allow this go to go through.

38:10 And in the meantime, let me copy our Okay, so here master stops the flow and it and it requires me as a human to approve this actual transfer. So as you can see here, this here's the circle CLI command. It's deterministic. It's showing what address I'm sending to the amount and the token USDC and what chain and a few other pieces of information here. So, you know, this is spending real money. So, I approve this and it will actually go through and spend the money. sorry.

38:49 Let me go ahead and copy our demo prompt here. And it was approved. And now it's sending the 0.1 cents to my wallet. And I'll tee up this prompt here for the next part of the demo. And as you can see here, it completed. It sent, you know, a fraction of a cent to my to my wallet. And it shows a network fee and all that. So, as you can see, it's a very, very small fee. Makes economic sense to send a fraction of a cent because look at my fee. It's like really, really small, almost free.

39:28 so we'll start our demo here. So this is the prompt here for you know my agent to do some you know research on the company right before my meeting. It'll make a bunch of different decisions on what how it spins things on agent marketplace. I'll I'll intentionally decline one of the actual decisions just so you can see how it kind of re kind of it just makes a different decision about like maybe another service it might want to use or cheaper service it might want to use to to get to the completion of the task. But here it's doing a bunch of things as you can see here. It's searching the mark it's using CLI to search the marketplace for relevant tools.

40:11 is searching different terms and we could even expand this open here just so you can see what the CLI command looks like. Circle services search and it's searching for these different terms so we can trace every single step. it's checking its balance to to make sure it can even afford some of these these tools here. it kind of gives me a status in between. It says it plans on using three complimentary services.

40:52 you know it said I'll estimate every call and keep each at below 5 cents USDC total spend below 15 cents which are our daily limit. So it's reasoning about that. It's found some services here that it's inspecting and it's it's checking the pricing and how much it needs to spend. here we go. So it wants to use Exo web service which web search which is a fraction of of USDC. and so it it showed only one. This time it only wanted one different one service. So I'll just approve that one.

41:29 Normally it does multiple. I did a previous run here so I could show you kind of what that looks like. And now it's paying for the actual service. Searching information. It's going to ask me in a second to approve that transaction. Oh, so it's actually looking to use a different service. Now it's using orthogonal. So now here it asked me.

42:06 Okay. it's asking me to it wants to pay 0.2 cents for a a search on orthogonal serp. so what I'll do is u maybe I'll just decline this one and see what happens. see what it decides to do. It says declined. It wanted to retry it. It's switching to a zerocost alternative. So, it's just going to go for a free one.

42:37 you know, sometimes it actually comes up with a few different services. It might come up with three or four different services and I could decline, you know, some of them or allow some of them and it, you know, it it'll make its own reasoning and try to get to the actual result. let's see here. So, I think it's about to finish up pretty soon. it paid for EXA, and it's great because you can see exactly every step and what it's doing here.

43:11 and you could also see this stuff in Master Studio as well, which you know, you can go to observability and traces. So, you can see all the traces that I had. This is the one I'm running right now. and it finally came up with a brief. so the brief came through with actual information from behind a a service that had proprietary information. it gives me a ledger of what it actually did. So this time it only used one service. It it only charged.7 cents for that service. the other ones are Oh, so here's the serer service here. I declined it, didn't spend anything. Then it found a few other services that were for free and it gives me kind of a breakdown of what it spent. Tells me what budget I have remaining. And then it actually sends me a link to the transaction that happened on chain. So I can actually see that, you know, the transaction actually happened for real on chain using USC 7 cents here.

44:14 what I'll do is I'll show an example here previously of when I declined another service it actually said it declined so and I I won't retry it. I found a lower cost more relevant alternative and so it actually went and reasoned and found another service even because I declined the the previous one. So this is a different run here. and as you can see here, it made several different payments to different services to get that information for the for the brief.

44:51 But yeah, so that's the demo. We've seen, you know, an agent, you know, spin up a wallet, you know, use its, you know, Circle Circle Marketplace to find services to, you know, help it complete a task. It spent the money autonomously. it allowed me as a human to approve those those transactions. If you want, you could also allow the agent to do it on its own without approval. Just setting those policies and parameters around the wallet to make sure nothing, you know, bad happens and then yeah, it autonomously makes its own decisions and send spends the money and then puts together and completes the task. So, yeah, that shows agents that can pay.

45:34 so I'll switch back to the slides right quick and I want to put up a few links here of how you all can get started with experimenting with some of these tools and things. I've put links here. the master template is not available yet, but it will be very very soon. working with Brandon on getting that open source so you guys can start playing with the stuff. But thanks. Thank you so much, Anthony. That was that was awesome. And this is the the kind of thing that I feel like we could have a three four hour demo about. We really just started to like scratch the surface here. But >> truly incredible. We we have a few minutes for questions here. feel free to throw any in the chat that you have. But one thing I'd love to get your perspective on and that I know I've heard a lot in calls I've had is, you know, obviously a lot of what we saw in the demo is like paying for these different services, sending it to people's wallets directly, a lot of sort of services related things. I'm curious if you have any thoughts on how we'll if how we'll sort of trend towards agentto agent payments and how you're sort of thinking of that space like directly going from certain agent wallets to other agent wallets. For the person in the chat that said DH is looking for you. I he's definitely not going to like this question very much.

46:58 But I'd love to see within Circle what you guys are thinking about in terms of like agent to agent versus a lot of what you showed today which is like sort of agents for services or things like that. >> Yeah. Like it's it's starting now with like agents to services cuz there's you know there's a lot of SAS companies and and capabilities online already that you know agents would love to use. So, it's great that X42 has come along and allows you know established companies that already have APIs to unlock a new business model that allows them to sell to agents directly, right? With capabilities that already been developed, they don't have to develop anything new. it just opens things up to where agent can come and just use that service and buy a piece of inference or whatever. but then like I once that matures more I feel like agent to agents is kind of like the ne next step right and when I think about agent to agent agent to like a agent may want a capability of a more advanced workflow rather than like a kind of a static workflow that a API has right like you can make a request API and it gives you back some data but what if you want like what if there's a specialized agent that has like some real specialized capabilities can also reason on behalf of you and you want to tap into that right as a agent. that you know I haven't seen we haven't seen a lot of examples of that yet but I think that's the next stage is like an agent sees another agent that has some capability that's proprietary that it wants access to. It's like hm maybe you can help me out with this right and >> send that send that task to that agent.

48:40 the agent does this proprietary thing and returns you know some data or even starts to work on behalf of that other agent right maybe it's a continuous task or something maybe it's you know you know like a financial trading agent or something like that th those that's actually an example that we see today is like some of the the trading agents and things or like financial to make to to you know invest you know money into different markets and things like that like a agent can actually make those decision and autonomously like trade and and make financial decisions on behalf of a human, but I do see that as a next stage. we're not fully there yet, but it's exciting.

49:23 >> Yeah, 100%. And I I think I mean for Moser internally, we're you know, all of us are kind of building our own internal agents here. And it'll be funny cuz we'll all build different ones out and then we'll be like, "Oh, my coworker actually built out this agent that I'd love to just like plug in rather than like rebuild that little section of it." So, you know, obviously we're all on the same team, so we're letting each other access each other's agents. But I can totally see how very soon that can turn, you know, when you're dealing with outside agents of being like, " you know, I could spin up my own agent that does this, but they do this really well.

49:56 Why don't I just contract out that agent to handle this part of it?" so yeah, really the sky's is the limit with with all this stuff. so it's very very exciting. I I'm conscious of time and I I know we're we're getting towards an hour. Okay, we'll do we'll Sebastian just threw in a question. So we'll we'll hit Sebastian's right a buzzer beater right before. Are there options to roll back payments if something goes wrong? >> oh that's a good question. you know, onchain payments are are permanent. So, you know, once that, you know, once that payment is signed and verified and and and transacted, it it's permanent. that's why it's so important to have these policies and controls around things. you know, as far as like, you know, if I guess the question maybe is around also like if if something happens and they need a refund, right? Like could that happen as well? that that can be that can be something that I think the the industry will evolve and and maybe support like refunds and things like that if a service returns a a a bad result or something like that, right? Like having something automatically kind of trigger a refund.

51:14 but I think the you know the ecosystem needs to evolve to that. But to answer the question, yeah, those payments are are are final on chain, right? It's a immutable ledger and once those things go through then they go through but it's important to have controls around that to make sure accidents don't happen. >> Okay. Awesome. And then and then one more we'll do we'll do this one as well. Are there any ecosystem discussions around integrating these payment protocols into MCP protocol?

51:46 >> integrating these into MCP. So so MCP is is really good for like it's a useful interface for exposing like tools to an agent. you know I think there's different layers to this like MCP is the this like kind of like the interface for exposing the tools. There's like the wallet stuff that we just went over with the policies and things. and then there's a reasoning layer, right? I think all these things kind of like go hand in hand. and yes, like yeah, all these things like all go in hand in hand and they all work together.

52:24 I'm trying to understand like the question of integration with Can you read the question back again? Sorry. >> Yeah. Yeah. Sorry, I'll I'll pop it back up here. Are there any ecosystem discussions around integrating these payment protocols into an MCP protocol? >> Yeah, I mean, Xro 2 is is is the payment conversation. So like if you have an MCP server and you want to like charge for that, you can put a XO2 in front of it.

52:50 Like it it really is a protocol that like I said of of having a conversation in a language of how to pay for a service. Basically it could be API, MCP, you know, anything. It could even be a a website, right? You can put X42 in front of a content on a website, have someone pay USDC to access that content on a website. really X42 is a protocol to like have that like payment conversation between two different software systems. So yes.

53:21 >> Yeah. And and I think a lot of that as well is sort of like what's the goal of you making this this payment agent, right? Like if you're basically developing a payment service or something like that and you want to expose it to as many users as possible like like Anthony said, you can definitely spin up to an MCP protocol. I mean, if you're making your own internal agent and you're just using it, then then it's sort of, you know, you're just using internally at that point. So, the the exposure isn't necessarily as important. and yeah, Sebastian, just looping back sort of to your answer.

53:49 Okay, double and triple testing. Got it. I mean, yeah, this is for any agentic system, I feel like the theme of of 2026 to a high degree now that we're sort of a little bit more used to agent development and the hype and everything it can do is having proper guard rails. So really that's a pattern with all these new things is you know it's the same as when people would occasionally build agents that would just fall into a loop of continuously calling an LLM over and over and over again. You got to make sure that you have guard rails in place to really prevent you know going over cost or token usage or anything like that. So, and luckily, you know, that's that's a great reason to do the circle and master together is because that's definitely like a core focus of both of our products. so you have all the tools you need to have to have, as we saw in the demo, you know, there's ton of human in the loop.

54:43 Circle, you know, has a ton of sort of last defenses of, you know, the different policies you can put in and all this stuff. But yeah, it's all about guard rails and, you know, we're also always happy to help out with that stuff. So feel free, you know, if anyone's ever building anything is interested, chat more about this. I'm popping up my meeting link on the screen. Feel free to grab time with me. I'd love to hear I mean I again I'm very very interested in this space and I would love to hear what people are building out there. so feel free to connect with me. and then before I'll I'll leave it to Anthony to close just to chat about, you know, maybe where people can find all this stuff and and if they want to reach out to his team or anything like that. But just to touch on the slide he has up here as well. yeah. So, Circle and MRA I I think is sort of a match made in heaven in a lot of ways. So, there's going to be a lot of stuff. This is just sort of the starting 101 course to bring a lot of you in, but we'll be sending a follow-up email this week that will have the template linked. So on Monster's website, you'll be able to just spin up a project that already has the Circle payment agent integrated and you can kind of get started there. So if you don't have any ideas on what you want to build yet, you can sort of start with a boiler plate and take it wherever you want. And then yeah, outside of that, we're we're definitely planning to do a lot of inerson stuff, you know, maybe hackathons building on this stuff. So, feel free to reach out to us and let us know, hey, if there's a city that you want us to try to come to and try to plan stuff about. definitely let us know. But with that, I'll I'll pass it to Anthony if there's any last things he wants to chat about for reaching out and contact.

56:23 >> yeah, definitely. definitely check out agents.circle.com. That's where all of our agent stack stuff is at. We'll be coming out with a lot, we're always updating that site. We're coming out with a lot of different tools around agent commerce. we'll be coming out with some circle products that get that kind of supercharge capabilities of agents to to do financial operations on chain. those tools will be coming out really soon as well. beyond just spending USDC is is going to be some other stuff.

56:54 for people that have APIs or capabilities, definitely go to our marketplace site. You can actually apply to like get your service. well you need to integrate nanop payments into your service but like you can get your service listed on our marketplace and agents will go and spend money there and actually pay you for your service. definitely check out our X channels circle and also ARC and also you know our developer X account. we have a lot of announcements there. I'm constantly releasing a lot of open source software apps and starter kits and things around agent commerce and other products that we have. and you know as I mentioned earlier the template for the master template will be coming out pretty soon. And also go to you know just continue going to our our ARC website. ARC is going to be launching on September 16th. So super excited about that. And we also have a community site on the Arc website as well. so you should go join that and you can interact with our with my team. we're developer marketing. So you know if you have really cool projects and things, share them on there and we'll we'll you know interact with you and you know potentially do some like you know web stream like a you know a live stream or a you know amplify your project or do some co-arketing things like that. So we love hearing about really new and innovative projects. I also run circle grants.

58:27 So also go for you know apply for a circle grant on you know our grants website. yeah and I think that's it. Just trying to give you know all the resources to developers. and yeah that's about it. >> Amazing. And I I threw up the link there to agents.circle.com. Obviously, there's a lot of other places you can go as well, but maybe that's a good starting point. so yeah, feel free to check that out. And then in this last minute, I want to give my sincere thanks to Anthony. An amazing demo, amazing presentation. really appreciate you going through this stuff. I know, >> it's a really, really valuable topic and I'm sure everyone in the chat is super pumped to be able to dive into this stuff now. I think it unlocks a lot of doors. Whether you're building a startup, whether you're at a large company trying to build different things, or you're just a solo developer and you want some cool agents that can do more stuff for you, this really opens the door to so much different stuff. Thanks everyone for coming today.

59:25 again, feel free to reach out to us if you have any questions. And with that, I'll let you all go to the rest of your day and hope you all have a great time building with Circle and MRA. See you. >> Thanks.