Credit for the Machine Age: Why DeFi's Missing Primitive Isn't Intelligence

Transcript

It's on. Hey. Hey. Thanks for coming. I'm Alex. I'm a sprinter um project focused on onchain credit and um yeah, great to be here. First time in Khan. Yeah, I thought um it will be cold and that I have to bring a winter jacket, but uh gladly it's not. Um yeah, I want to talk today about credit. um and credit there's a lot of uh in a lot of ways that's being used like private credit or different forms of lending but I what I want to focus on is something that I call um programmable credit and actually how it can be used specifically in the context of AI agents um to economically empower them or scale

actually the machine age um but I want to start more on that side uh where we are today so when you look at onchain usage then um it's kind of obvious is that for quite some time most of the onchain usage already has been um automated software. So um when you can see here that 86% of the trading volume is already generated by bots that might be arbitrage bots um marketmaking bots me stuff like this. Um so that's already the majority of the trading activity. At the same time there's something called solvers. So when you have for example interacted with cow swap or with across or another type of intentbased protocol then you have been leveraging solvers to

actually um carry out what you want to do for example swaps or crosschain bridging and stuff like this and they also do like solo based protocols already 10 billion and more than 10 billion in monthly volume actually. [clears throat] Uh, and last but not least, I want to touch on um AI agent volume. Um, and I think the top AI agent that I could find that currently on on one of the top marketplaces that actually offers um swaps and basic things such as DCA, so dollar cost averaging and um and also trend analysis for tokens has done almost 200 million in total lifetime um volume. So that's yeah almost not as in the same order of magnitude than um bots and other automated software but um it's kind of

clear that AI agents are taking on and and growing and that um starting in commerce in agendic commerce um for example they are um yeah they are basically starting to take off. Um, next I want to touch a little bit on autonomy because when we talk about AI agents and bots then you could ask yeah what is the main thing differentiating them and that is that in the AI agent there's an LLM that means a model such as chat GPT or OPOS or codeex and that's part of um basically the execution path of the software that means um it basically takes decisions as part of the execution of the software and and and acts on top of them. What you can see here is a kind of framework I think was proposed by the

Columbia University New York that actually shows a little bit the different stages of basically autonomy and and what I think first and foremost this should I this should um show is that autonomy is a spectrum. So um it starts more on the left side and you see a level one for example where you have more or less the human still as an operator and um the human actually making most of the decisions and just consulting the model to provide input or research or whatever and then it goes all the way up to level five where actually the human is more or less only the the observer and actually the model drives most of the execution and decisions. Um so obviously user involvement uh or human involvement goes down with the levels and then the agent autonom autonomy increases. And when we

think now about um about what we see today then we have to kind of attest that um this the level of like where agents are at the moment really differs on the domain that we are looking at. So when you think about something like economic activity that's what we are mostly interested in that means onchain economic activity then I would say most of the activity is probably here or here that means like a lot of the bots for example the bots that we are running for the most part for liquidations market making and stuff like this they are still fully deterministic in the sense of there's no LLM in the execution path the LM might have coded the bot and contributed to that but um the execution path itself um actually for economic activity is still very much on the left

side and then that obviously differs for for other domains. When you think for example about coding itself that means actually um um yeah software engineering then there obviously the agents like claude or codeex they already very much on the right side of that spectrum they are almost level four level five and you can see that here um shown by this example from entropic research which actually shows how the turn time that means the time that um claude is basically working without human interruption um this turn time has almost doubled to I think something like 45 minutes just over the span of the last four to six months. So what you can see here that yeah it's very different as said on on depending on the domain

and in coding for example um we already are at the point that um the agents are more or less autonomous and require very little guidance. Um yeah, as a credit protocol, we also obviously want to look at something called guardra because autonomy um obviously requires some kind of guard race. It means obviously um it's good when when the model can can act on its own um yeah on its own determination basically. But um for reasons of performance and reasons of of of security, we actually want to be able to kind of um provide a harness to the to to the model itself. So that's why you could say an AI agent something like

yeah cloud for example that most of you might be familiar with or codeex is actually a model in that case that's um that's maybe um oo or sonet and then a harness and a harness is then a set of tools or a set of um basically guardrails that the model that the model uses and they it use a set them for two main reasons being performance that means it has the right information and tools available to do the job you know in the case of a coding agent like cloud that would be it has access to your um to your CLI it has access to your shell it has access um to your code repositories and it can actually make co make code changes it has the required credentials and stuff like this but at the same time obviously um there's also

a security relevance here that means obviously we also want to ensure that only the required credentials or information um for the model to do its job are there we as a credit um yeah protocol are obviously interested in at economic side of things. So more and more we see those um agents when when they are then um acting economically to also have an economic layer in that harness. And that means there's things such as payment methods. So they might have wallets or virtual debit cards that they are using. They also might have um um integration with payment rows. So there's agent SDKs that then offer integrations with the measion payments protocol, the agent commerce protocol, um a stripe checkout when there should be a human in the loop and stuff like this. And obviously also what you can see in most of the economic um agent

SDKs is things such as spending limits or maybe merchant categories to actually yeah kind of guard or limit what the agent can actually spend the money on. Yeah, for us um from a credit point of view this it's now an interesting question. Can we do actually something um with those guard rails um to actually go beyond um simple preunding of agent wallets. So the current prevalent mechanism to actually um fund an AI agent if it should act economically is we give it a wallet and we actually um issue some or transfer some stable coins to that wallet or some some other types of funds and let um the agent run with it. From our point of view obviously you are credit. So we would ask why do we

really need to fund the agent at this point? Can't we just actually issue credit just in time when the agent really needs the funds instead of having it to hold the the funds all the time? Because this is obviously not good from a capital efficiency point of view, but also won't scale to millions of agents that we hopefully see soon. Um the second thing and that's the interesting part is like um usually when we think then about credit and using credit with agents then usually in the onchain realm that means we have no KYC available. We have no legal enforcement or offchain agreements. How can we actually yeah offer good terms for those agents? In current lendings, the lending pure permissionless onchain lending usually

the only way to go to credit and to even certain forms of cheaper credit is collateralization. That means providing um actually um yeah providing collateral and then borrowing against that collateral. so that we as the creditor we have it actually an enforcable claim on the on the collateral and can actually um liquidate in case of debt situations and stuff like this. So the question is now can we actually do better because in the agent situation there is definitely no KYC available. So there's definitely no credit scores or offchain loan agreements available. So ideally we want to be able to issue credit but at the same time in a permissionless and KYC free fashion and in a capital efficient fashion. So that would be ideal. So let's see um if we can actually go there and that by that by using pure onchain mechanisms you

know everything that is there without the need for set KYC and stuff like this and then even further can we actually use the agent harness it means the methods that are now available and spreading like CLI tools APIs and MC MCPS to actually integrate those forms of credit directly into the agent harness so the agent can access it when it needs it and on demand. Um yeah, I want to now take a look at three examples because I think it yeah that's a little bit theoretical. So let's look at three concrete examples on how this could actually look like if we provide agent with um with actually economic access to to um cheap capital.

First one um crypto debit cards. Yeah, I mentioned already that currently the most prevalent way to uh to fund agent is actually agent wallets. There's another way and that's virtual debit cards that are being used. Um yeah, I think virtual crypto debit crypto debit cards in general did something like 10 billion in annualized spend volume last year. So it's um yeah, it's quite a huge volume already. And yeah, I want to now take a look at how we can actually use the more efficiently in in um yeah together with agents. So on the left side you see a little bit like in this box um how a credit market would have to be structured. That means it would issue USDC and stuff like this. But we will go over that in a little bit detail with the scenario. So in this case here um the scenario is basically that we have a

human operator and this human operator wants to actually um set up an agent you know agent should for example um um search and and and acquire and purchase gifts for a wedding maybe or maybe something um birthday um yeah and and once I set up this agent with a dedicated virtual debit card and and and um yeah for it to act um and this at at the moment I asset requires manual preunding. That means I can issue that credit card in in the in the case of crypto cards. There's usually an onchain address that is associated with the um uh with the visa master card. That means where the funds are being hold that then um is is basically the spend requests

are running against and um yeah how can we get around this actually measure of preunding because it's and make it more really credit card type feeling instead of yeah this debit card feeling. So what we can actually do now is two things and I mentioned this enforcable constraints. So two things you will see them a little bit like um also repeating repeating in the next examples that we think are really um instrumental to provide sorry to provide better terms um um in a KYCless fashion are credit accounts and credit operators and credit accounts basically means we can actually leverage the fact that we can issue those funds for the agent but put put them in a kinds of a kind of constraint box so

that the agent isn't free to do anything with them. No, so they he is basically limited in what he can do with those funds. Um so here in this case we would have to opt it would be still an overolateralized credit um credit line that means the human operator actually um has collateral on chain that backs the credit line and there's a delegated credit account that the agent is agent is using and um this agent this agent account is actually limited then to be used for cult spend. When we do that, we can actually ensure that by using a credit operator, which is actually the same as kind of like a liquidation board you can think of, right? That acts actually on the on behalf of our credit protocol and basically ensures that certain constraints and conditions are

met. Um so and these credit operators actually then with an integration with the card program is basically able to on the card spend request that comes from the visa network or from the master card network to draw directly credit from the agent's account on behalf and get funds access to funds directly via that credit line. Um what this gives us is basically the agent has a wallet. Um the agent doesn't need funds on that wallet. It basically can um access funds on demand. We need an integration with the card program network for this. But once this is available, we can actually um basically directly react on the spend request from the card network that comes when you want to swipe and use the card and then or when when there's a virtual

in in the e-commerce case. Um yeah, like a virtual checkout process for example and then spend against this. Yeah. What this also gives us is the ability to actually provide more free credit. That means um when you know like how traditional credit cards works work um then there are the fees or it's basically free of charge for the user. The fees are actually paid by the merchant. It's called ex fees interchange fees and it's basically that the merchant um who who who basically yeah sells something that he pays a fee and this fee is then split amongst the card network and some other parties.

What we now can do actually in this case when we know that our credit is only being spent because we have this credit account that sets those constraint we know it's only being spent for actually card spend on this card network then we can actually also have an offchain agreement in place and actually move our fees also to a share of the EX fee what this means is we can actually provide credit free of charge for users and in the same way like in a traditional credit card actually use the ex fees to to fund this. Yeah. Yeah. So, TLDDR it's a little bit complicated um I realize but um TLDDRs there's a constraint account um and when we use this constraint account and can be sure that the the the funds are only being used for card spend we can actually provide a more creditlike experience by actually

drawing funds just in time when they are needed and secondly by having those um spent free of charge for users because we can use the ex fees to subsidize this okay >> subsidize >> um it's always risk yes I mean credit price is risk of loss and the exe would then price price that. Yes. Um okay let's now go another example. This um like the first example was still over collateralized that means we still need more value in collateral then um we can issue in credit. Now we talk about leverage account and in leverage account u we actually start to go more into the underolateralized direction. That means the agent can actually access more funds

than we have in collateral to back that up. Um in this case we have a trading agent. So we saw that initially so 86% of the trading activity is already um is already done by bots and uh increasingly there will be AI be part of those execution paths um and and do parts of the of the trading strategies. So let's assume we have trading a trading agent and this um um he has identified um or the agent has identified a profitable opportunity on a decentralized exchange and wants to increase exposure while um having limited funds. It means agents has make maybe 5k worth of USDC to spend but wants to actually maximize

exposure because he knows the agent knows returns will make up for that. So what we can now do and again you will see here the example of the credit account and the credit operator we can actually issue more credit than we have in collateral. So assume the agent gives us those 5k okay those 5k in in in collateral um we can actually issue 10k um in credit to this limited credit account but we apply very strict guardrails on this account. So the agent can't simply move the funds out. the account is actually um is actually customized for the venue to trade on. So it might be for example hyperlquid, it might be unis swap and then the agent can actually execute um certain actions

on with those funds. So he can do things like open positions, spot perpetuals, whatever. But at the same time there's again a a credit operator in place that from our behalf from the protocol basically ensures that um um yeah that the funds can't that the balance always justifies all our margin requirement and um as soon as the margin requirement gets breached we would um liquidate this positions. I have to hurry a bit. But DLDDR on this one is like with applying even more constraints, we can actually give an agent in a KYCless and permissionless fashion access to 2x leverage. But for that, we have to put the type constraints very tight. We have to basically limit the target venue that

it's being used on. you have to limit the actions that you can use on the target venue and then we can actually give you 2x leverage without KYC and permission set and at the same time be sure we will always always be made whole and our margin requirements can't be breached um went oh sorry um yeah was way too slow um yeah so we can't go through a third example I will only shortly touch on that so last but not least solver credit we touched on solvers already when you when you kind of take this example to the to the extreme. You can actually provide zero collateral credit and that's what we and sprinter have been doing um since half a year roughly.

So um we have done I think 300 million in credit volume using this approach and it provides actually solvers that means entities automated software that wants to for example solve in across in Mayan or in other crosschain protocols we can provide them without KYC permissionless access to zero collateral credit and at the same time be sure yeah that we always be will be repaid when you're interested in how that works check out our website or come to me after the talk because I think I'm at time but I want definitely use opportunity al to also um yeah kind of share miso with you. I'm not sure if there's any crypto debit card users in here. Can I see a hand maybe? Yes.

So if you do um yeah I mentioned initially already a little bit the challenges with crypto debit cards in terms of in terms of managing the cards, topping them up and stuff like this. So, if you want to level up your experience, how you um earn, manage, and top up those cards, make sure to follow us. And also um yeah, find us here around the venue. You can identify us with the sprinter logo and then yeah, we might be able to tell you a secret or you get some early access. All right. Thank you. [applause] >> Thank you.