The Infrastructure Behind Agentic Finance | Roundup

Transcript

these kinds of things around like agents working on behalf of us and having stake behind them. I think is the next era for crypto and AI is still super early and to be honest there aren't many teams actually working on this in the entire space but for me it's kind of clear we're going in this direction. Hey everyone, quick disclaimer before we get into today's episode. Nothing said on bell curve is a recommendation to buy or sell securities or tokens. This podcast is forformational purposes only and the views expressed by anyone on the show are solely our opinions, not financial advice. Our guest and I may hold positions in the companies, funds, or projects discussed.

>> All right, everyone. Welcome back to another bell curve. You got me, Miles, and Zave. Miles, are you socked in there in the East Coast? >> Yep. I'm buried. We've got like 3 ft of snow outside. Um, it's like arctic weather and it's kind of funny. Uh, you're in Montana hoping to go skiing and there's no snow. I'm in Boston hoping to just get around easily and there's tons of snow and we can't do that. So, all good. We just >> I picture you kind of getting out there doing snow angels making snow.

>> No, I mean I did. I did after the Patriots won on Sunday, you know, going back to the Super Bowl, City of Champions. >> Afraid to jump back on the bandwagon for me. This is >> You were the most pathetic Boston sports fan I've ever met and I love that about you. It's very endearing. >> I feel like I lucked out and basically the best luck a sports fan could ever ask for is like slapping God in the face for giving you a beautiful gift just not just wasted on you >> and you just don't care. [laughter] >> No, I support our boys.

>> You do. You do. You do. >> Yeah. Um All right. Uh okay, guys. We're going to jump back into uh so this is actually the third really episode that we're doing on uh the intersection of AI and crypto as a rehash. We talked about the evolution of open source um and really software in general and we talked a little bit about how different corporates um leverage open source as a business tactic and we ended with AI and crypto um and how both of these evolutions in software are really changing the game. And we did our first episode of uh AI and crypto on where are the intersections that really make sense and one big category that we talked about was deepen um and what we could learn there. We wanted to dedicate an

entire episode and probably close this series out with an an episode dedicated to Agentic Finance, which is a which is a crazy big buzzword at the moment. And um not to denigrate, but anyway, Zave, you you're kind of our resident expert here. So, I'm going to kick this over to you. Introduce us to the concept of Agentic Rails and why people are so excited about this particular intersection between AI and crypto. >> Yeah, of course. Thanks for the intro, Mike. And yeah, hello everyone. So, yeah, this episode I've been looking forward to. I think everyone in the crypto space is quite excited about the intersection of crypto and AI, whereas one might also make the argument that

maybe it hasn't lived up to its potential yet. Uh there's plenty of standards being worked on which I'm sure we'll crack into this episode like MCP, X42, uh etc. And there's very likely that these standards will merge. And so realistically right now the infra is being built. So yeah, we'll touch upon what infra that is, why we need it in crypto um and where that is at right now. And I think in particular because the AI space is moving so fast right now. I do think a lot of the opportunity in terms of startups and what to build in crypto that makes the most sense is actually at the more kind of application layer and really using these rails in unique ways to target new markets. And so I'm sure we'll talk about some of those opportunities. But I think in general it's a really exciting time in

the crypto and AI space. It still seems very very early. I don't think anyone has very very clear views on exactly where value will acrue, what will be needed to push the space forward, etc. I have some opinions I'm sure I'll share in this podcast. But in general, I think the exciting bit about crypton AI is sort of the openness of how you use the two technologies as opposed to the closeness that we see right now with some of the major players like OpenAI. And then the question is there like at what point will this actually come about when this open ecosystem really flourishes and emerges? I think it will probably take several years to build out. But at the same time, Open AI, these kinds of companies, the infrastructure is kind of commoditizing.

So yeah, it'd be interesting to see I guess as that commoditizes where that value then acrews the open source stack. >> Can we talk about before we even get into the specifics here and Z I want to get into X42 and some of these different standards that are getting developed. Can one thing that's never been super clear to me even in before the research in this episode of why is why do payments and finance on crypto work better for AI agents than they would in the legacy system, right? like is this the kind of thing where uh two-factor doesn't really exist for the most part yet in crypto? So it's just easier uh you know what what is what is the reason right why it's easier for agents to transact on crypto rails as opposed to just like

hitting your banking API what do you guys think >> yeah I'd say right now it really simply traditional rails have been built for humans and so you know as we all know banks have closing hours a lot of these ecosystems are closed you need to pay for APIs it's it's really cumbersome to work across systems and this was built for a previous age which we're not in anymore. We're in the age of kind of AI and crypto and you know really micro payments etc. And so the opportunity really in crypto AI is about making this a lot less cumbersome than what it has been in the past to have open rails that run and settle 247 micro payments with low low transaction fees. It's about

doing things that weren't possible before in this environment where everything is always on and easy to be accessed and easy to be used and paid for. So to make all of our daily lives better. That's what I would say in a nutshell is like really what it's it's there for in the first place. And some of those rails would touch on are these standards that are being built right now. >> Yeah. Like like on the surface there's like a pretty obvious uh like marriage or really good fit here. um as we I think we have basic it's kind of consensus right at this point that like AI is going to transform a lot of what we do today um and it's already like most you know seen we're already seeing it in in terms of programming right so like even the best developers in the

world are very happy for AI to write pretty much all of their code and now they're just doing reviews right um and we it's it's unclear that that like step change happened very fast that happened in like the last six months only right um and we it's pretty clear at some point in the next like 6 months to 10 years that AI is going to be you know trusted with moving money right and we're going to automate tasks that used to be performed by humans we're going to have like different interfaces right where we're now talking to like um you know an LLM about what we want to do and then it's going to we're going to fund an agent to go do that um and it's just not clear like how soon that's going to

happen. But yeah, the reasons are obvious like we these things agents run 24/7, crypto runs 24/7. Um we need the ability to like lower the cost of payments if we're going to have you know agents making microp payments for something like you know per call APIs and things like that. Um, and I think yeah, the bottleneck just becomes like trust and how can you ensure that like how can you verify basically in the same way that you're doing code reviews, how can you like automate verification of something like policies and guard rails for like these agents. Um, and frankly, it's a lot less scary to put like $20 into a crypto wallet and give that agent permission to spend than it is to give

the agent, you know, your bank account credentials or your credit card credentials and hope that they don't, you know, overspend, right? Um, and we can program a lot of these like, you know, uh, limits and policies as as smart contracts. Sometimes easily, sometimes, you know, it's more challenging, but that's the general idea. Yeah, I would maybe add actually as well, Miles, is that it's trust is critical of course, but in terms of the value that we get as humans at least right now, I think it's trust across very complex workflows because there's something to be said about having trust around you know some policies or permissions to make one transaction on one blockchain. But then the question is there like why would we need an agent to do something as simple as that? And I

think that's why agents haven't seen a lot of PMF right now in crypto because a lot of what they're collapsing as workflows are very very simple and as humans we can just bypass that into something that we already are used to using that we already trust that already works for us and so I think it's actually trust that is the constraint as workflows get more complex but that's also actually where the value is I think for users right now in crypto which hasn't been explored enough. That's super well said and I think I mean one thing that you know it's been interest maybe there's an analogy here around you know as we finally have institutions moving into and building on top of crypto what what value do they really find and it's obviously there's a predicate layer of like decentralization and neutrality and

all that stuff but really it's it's 247 right I mean that's such a big unlock and advantage from how it works in finance and I can very much imagine you know if you're imagining the UX of hey I my agent to do something, you're dumping something into a chatbot getting returned a bunch of error messages. I mean, just literally look at from the perspective of a product manager like, hey, I want to enable people to use an agent and if you're highly limited by the rails that you're operating on, nothing to do with the LLM, that's a massive constraint for your product. Um I guess where where we talked about maybe there's a an element here of um path dependency and timeline and I would

love to get a sense from the two of you because the the question that I still have is how long is it going to take to build out those uh permissions and guidelines and trust because right now it's just the the trust issues. So on on the one hand, I very much see the potential, but I also I'm I have so much, you know, I've been burned before operating on chain and the thought of adding an LLM there, it just the the monkey part of my brain's like I don't want any any part of that. So I I think it think it's like how do you get from here to there? And one thing maybe we could do is start talking about what some of these standards are trying to solve. And in particular, I would love

to drive towards Miles, how some of these old episodes that you and I did on intents and account abstraction and things like that, like how that's suddenly extremely relevant again. So maybe we can maybe we can start by talking about some of these standards like X42, uh, you know, and the standard that I forget the name that, um, Stripe and Shopify and opening up. >> There's there's there's so many. Um, and yeah, Ze, feel free to like jump in here if you feel like you have a better handle over them. Um, but yeah, maybe we can start up like on on X42, which it was created by Coinbase, but it is actually not. Uh, it's it's pretty agnostic to the to the payment rail. Um,

but it basically just standardizes payment for access um on for any service that's provided over the web. Um, so it if you are calling let's say an API, right? Um, and maybe it's like the Blockworks data API that can now be accessed by agents because you have some sort of like pay-peruse, you know, uh, pricing for them. The server will return like a what we call like a 402 message which says like, you know, failed because you need to pay, right? And then it allows very easily like the the client so like the consumer to retry with a signed payment. Um, and so you can imagine this, like let's say I'm building an application or building some tool or feature. Um, I'm working with

Claude and I'm now going to give it, you know, $20 of stable coins. Um, and Claude's going to call a bunch of different agents for different like tasks and fun and functions and a lot of those are going to require API keys. Um and today there's a bunch of friction around like the user has to go in to that site, buy the service, get the API keys, give that to cloud, right? Um and the idea is the you know over time like the uh API providers are going to have like um basically a choice if they want to disrupt themselves and um make themselves available to like you know this more like agent native you know

payment setting, right? Um, and so you'll fund an agent, it'll go task a bunch of agents with doing things and then you'll, you know, basically have like reduce that that level of friction. Um, so like that's that's probably the the biggest one. And then there's a bunch that are doing similar things. So I think it's yeah, UCP um or ACP, I can't remember. That's the OpenAI and Stripe one. Um, and some of these are I won't go through all of them. They all have different like you know uh abbreviated names >> X42 >> A2A yeah some of them are for communications some of them are for payments um and they're you know all generally like working towards the same future but Ze Yeah, please correct me

for for what I might have missed. >> Yeah, I think very well explained Miles. There's a few mentioned or that we probably should mention uh for listeners. So one standard in particular which works pretty um nicely with uh X42 is MCP. So this is model uh model context provision or like I can't remember what the P stands for but essentially it gives a model like an LLM context over what you're trying to do and it's like a connector like a USBC port where like if you're using something like claude or claude code or something like that it gives one standard way of accessing any data across all tools from all the way from APIs to your local stuff on your PC. So

you can imagine intersections here where you're using something like Anthropic and you're trying to access services via API. And then this is actually where X42 comes in very handy because you can actually pay for things much cheaper, much faster and also in smaller amounts. And and Mike, we spoke about last podcast, you know, your own business with Block Works right now and sort of like usage based pricing versus per seat. I think really X42 is the early innings of what this might look like of what it's going to look like to use the gym one equipment at the gym and pay $1 instead of paying like a hundred bucks when I go once a month because I'm lazy and I'm not really getting the value for my money there. That's the analogy I think here, but it's going to become onchain. Then outside of that, more

relevant things to crypto are kind of around like ERC8004, which is agent ID on Ethereum. This is kind of early right now, but it's a way for us to trust agents further to know what agents we're interacting with. And there are some other things that are also like crypto adjacent. So, you have something called UDP, which is being worked on by Google, which is a bit similar to the ZKT TLS primitive uh in crypto that we have today where we can kind of have data offchain and prove that onchain and then build services around that. only this time it's actually being built by Google and some other people around actually what data we can bring to any service not just crypto but also to models themselves and verify that in a different way to how we use blockchain today. So I think really here the opportunity is actually mixing

these protocols because there's so much you can do and so much being built and you can find really creative ways of serving new markets with these protocols. So it'll be interesting to see as these microp payments come out, reputation for agents comes out, data ownership and sharing comes out uh everything from communication to connection etc. like how we plug and play these things in crypto I think will be huge in terms of where the opportunity might lie. >> I um I'm going to keep going to bat against the two of you here that unbundling everything is something that people will choose to do. I I firmly I'm actually like I'm kind of a hater of the microp payment thesis in general but Miles is like >> I don't know I don't know if it's necessarily micro payments. It's more like uh is there going to be another

version of SEO, right? Like but it's more, you know, >> totally totally >> and like if you look at like Superbase is a great example of like [clears throat] which which touches on uh a theme we talked about before. It's like a commercial open source software company. They're like the go-to database for um coding agents uh because it's so easy to onboard and it's so cheap, right? Um and they've grown like 3 to 4x in the past couple months. That's it. Like uh and because they've won this like emergent sort of demand funnel, right? Um that I think we'll see like a lot more of. But yeah, jury is still out. I agree. We're not >> I um I don't want to distract us too

much. I I think it's a super interesting question whether or not the what LLMs are a new level of aggregation, right? In the same way that the internet was and social media was. And it's just f it's I'm I will be so curious to see how this one plays out versus the internet because we now have an example of the internet. So when Google started to crawl every media outlet like they didn't really know what was going to happen there. Like some people were nervous about it but it wasn't clear just how badly they were going to get smoked on that. And you know what what I would think would happen in the age of LLMs is that people h people remember that lesson, but it doesn't act. It seems like the tool is just so freaking useful that uh I I see people kind of connecting

everything to LLMs again. So I honestly it's so difficult to understand. Um it's it's just I find it a fascinating question to see what what's going to if we're going to learn the lesson of the internet or not. Um, but I I also think that uh what LLMs do really well is I I think the the money to be made here is reducing the amount of time that orgs can make good decisions in. Like time to decision is so important in an organization. Um I actually think that's the massive advantage that um well structured data and LLMs can do. Okay. But I want to redirect us here um towards what some of these standards do. And one question that I had for the two of you.

So um I' I've actually got to compare and contrast because I think it's pretty interesting to go through what are the different challenges for Gentic Payments and commerce in the legacy system versus in crypto. And it's pretty interesting. They're like very polar opposite to one another. But before we get there, I would like to ask I mean this is the most classic thing ever in crypto. Everyone's launching a standard. You get a standard and you get and it's all public good. They're all just doing it out of the goodness of their hearts. But we talked about in the commercial open source episode that open source software episode that generally open source software and setting standards has a strategic business value to it. So I'm curious like you know if it was really just a good for everyone

to use, why isn't Stripe like hey thanks Coinbase for launching X42. Let's just use your standard. Everyone seems to be competing on launching the winning standard. So what does that get them? Why are all these organizations doing this? [snorts] I mean, I think it's pretty obvious like you can uh it goes back to what I was just saying before where it's like this is there's this new distribution channel, right? Um to to monetize software. Um and if you are the winning sort of standard, right? It's similar to what I was just saying with Superbase.

is like, "Okay, so you've given an agent some money. They're going to go choose like the standard that, you know, is kind of like the default for the industry, and then they're going to go pick the like tools around that standard. And if you are the one to have developed the standard, you're definitely going to have the leading position of like the tools that are used that you can actually monetize around it, right?" Um, and yeah, I think like that is the sort of like obvious answer. Um, we've also seen that when it's obvious that there should be a standard and that you can monetize things around that standard, we'll have standard fragmentation, which is like the kryptonite of of, you know, um, sort of these initiatives. Um, but

yeah, say, please add any thoughts I might have not covered there. If you push, >> I think these standards are kind of similar to like the blockchainer ones back in the day where you had all these blockchain ones going live and there was a bit of competition, a bit of collaboration, value was acrewing in different ways, but everyone was trying to be the underlying infrastructure. Like simple as that. And I think you're seeing the same thing right now with these protocols and AI. Like this is the HTTP moment for the web back in the day, but now for AI AI. So I think the opportunity is large if you can be that infrastructure provider and then build services on top like we were just talking about. But there is an interesting intersection here which is actually similar to crypto as well of like open versus closed and so closed

ecosystems like your sort of open AI etc you know and stripe they're trying to build these closed standards uh and these are going to be harder for us to interact with across services for example whereas you have people like Coinbase X42 which are pushing much further towards these open standards and again it takes me back to my point earlier which I said around the value acrewing to workflows that are complex and some of the most complex workflows we have today across various tools and softwares not just within one ecosystem or open AAI. So I think the jury is still out right now with regards to like what tool will be like the new HTTP or if there's going to be multiple right now each of them actually does serve a different purpose. So there are reasons people are building these different standards and again like I do think that

they'll be combined in different ways and people will build products on top but I would say there isn't any obvious one that is like really doing the best. I think X42 now in crypto at least is doing a fantastic job with mind share and adoption and a lot of my best friends who are developers in the space who have been in the space a long time. The thing they're most excited about right now is X42 and what you can build on top of that because you can actually reimagine new products, businesses, services, etc. The other one I think which is not going to go away is we're seeing now with Claude Code, Claude Cowwork, Claude Bot, all this sort of stuff is MCP. It's like we have these models, we're going to use these models.

How can we get as much data into these models as possible? give them the context we need to actually improve our lives and like get to what we actually want to achieve with our with our common goals. The ones which I'm actually unsure about right now are the closed standards. And you also see in the web by the way that like the closed standards didn't win out. The open standards won out. I would probably put money on right now the open standards being the one that went win out. For example, X42, MTP, etc. The closed door standards is TBD. We're going to see with Stripe and Tempo blockchain coming out like is that going to use their standard for agentic payments? Like maybe. But then the question is who's going to adopt that? it's going to be much harder to make the case to adopt

the closed source stack there. So, I think jury's out on winners and it makes a lot of sense. It's a trillion dollar opportunity. You can influence the entire AI industry if if you win the game and I think that's why Coinbase is spending a lot of time there outside of just their generic crypto uh business. So, yeah, it's exciting, I think, to see all of it go live and the innovation is really healthy and and just to finish, agents aren't great yet and you need these standards to really make them great. So yeah, I think it's exciting because we're probably a year or two away from them finalizing the winners and that's when the agentic ecosystem will really flourish. >> Okay, last question on this and then I want to get to we've been talking about agents and commerce as one amorphous blob, but I want to get a we have very

different ideas of and there are different categories of commerce right around what agents can do. Some is payments. So hey, I have this intent that I want satisfied. I want to book a flight from Bosezeman to New York. Um, another one would be, hey, I want you to implement this trading strategy and trade on my behalf and maybe even generate alpha. Um, so last thing before we get there, I would love to talk about just the different mechanisms that um we are especially folks in crypto are thinking about um about managing that complexity. And Miles, I know you have more thoughts here, but you know, one approach would be creating a like very strict

permissioning around what these agents can and cannot do, right? So you um but there's there's an enormous amount of complex to it. The the other thing actually and this maybe is getting towards what we introduced this episode with which is there are lessons here about how crypto intents there are similarities to how intents work in crypto today with account abstraction and actually instead of creating a bunch of complicated if then statements you could actually use collateral. So if you look at the way that um like a unis swap x intentbased scheme works is the builder gets to perform the action and receive a fee but they have to post collateral before they can you know get that back. So, you know, collateral versus, you know, data permissions and all that stuff, like how's this stuff

going to be managed? >> Yeah. So, maybe like to to zoom out um and go back to that original like question of is it actually consensus that agents will use crypto rails? Um so, the benefits of crypto for agents is that it's digitally native 247 everything we talked about, right? The the costs are much lower. Um the the downside is that transactions are irreversible, right? Um and the legacy rails kind of have the the inverse, right? Uh considerations where like transactions are reversible, especially if there's fraud or like chargebacks or like any anything, right, that could have gone wrong like it's not it doesn't settle immediately. Um but you know,

that comes with a bunch of other frictions and like downsides. Um, and so like the problem I think with seeing or like the big challenge of um rolling out this like agents that can handle money safely is that agents are non-deterministic. Like >> they you you you type in the same prompt, you know, five times and you get back five different things. Um, you could do what I's doing, which is trying to basically make like LLMs deterministic. Um, and and make that, you know, verifiable in the future potentially. Um, but like they are manipulable. Um, they can get tripped up on prompt like injections. Um, we've seen this already many times. um outside of like the this this question of like

money, but it's more like leaking information and like giving people credentials that they just shouldn't be giving them to. Um and so the problem is like we want kind of smart contract like guarantees, right? You want to know that there's hard guarantees around what an agent can and can't do. And it's actually not very easy to implement that on blockchains. Um, so >> what the the guardrails that you can sort of implement are all fully trusted. Like they're all behind a box at the moment. Um, >> if you try to implement them as smart contracts, let's say it's like spend policies, um, or even like much more granular policies, uh, you can do that in some cases, but the challenge now is like that is

public. It's deployed on chain. So that in itself is a vulnerability. if there's a bug in it, none of those guardrails matter. If what the data that like the guardrails that are being publicly verified, you know, are out in the open, that's now an attack vector, right? Um, and so it I think the devil is in the details here of like how challenging it actually is to encode guard rails onto these systems when you're giving an agent that's, you know, fully non-deterministic has to get, you know, external data piped into it either via like ZKTLS or via some Oracle system.

Um, and that's what the teams that I'm talking to today are dealing with. Um, and that's what, you know, we're trying to like basically figure out solutions for. Um, and it's kind of ironic like we have identity is like really being pushed here, agent identity. So like there's a big ERC push around like creating a registry of identity for agents. Um, even insurance, right? Like this is goes back to like IEN's why I might be interested in this. Um, you want to have like if you're ask an agent to do something that is putting your money at risk, they should have something at stake too, right? Um, and I am just a little bit skeptical that we might need something stronger than like the agent's incentivized to behave or like the agent has a high reputation

score. Um, I I think these things are like decent interim options, but uh, as soon as an agent with a great reputation score, it's incentivized to behave up, you know, we're we're back to square one. Um, so that's my push back on kind of like what folks or like what what folks in traditional blockchains are are doing these days. >> You know, you know where that argument really resonates to connect that to another theme in cryptos. Just looking at lending, right? For years we've been talking about, hey, assetbacked lending is not how it's done in crypto. We have to move to uncolateralized. And it just hasn't. And I think that the superpower of crypto is the ability to move

collateral to back up a claim or promise or ensure something with such little friction that that should actually be the guarantee. And honestly, that's the guarantee that people want, right? Like, okay, my agent hasn't messed up in the last 50 transactions. Like, that doesn't do a lot for me. Hey, you have 100% guarantee because if the agent messes up, they've posted this collateral. And then it's a much stronger incentive. you're pushing the incentive to the edges for developers to figure out, hey, if I've actually in order to, you know, clip this fee on booking Miles's next airline or trading his uh >> right, >> sheu or whatever, >> whatever degenerate he's trading, then that's that's just a way better

guarantee right off the bat. And it it it aligns with what crypto does super super well. So that's kind of the way that if I had to guess, I I would see it going in that direction. I don't think like reputation, scored management's good enough here either. >> Yeah, I think great point. Uh we work closely with the team right now, which I think is actually quite cool and moving in a direction where I could see the space going, which is more focused on the staking kind of side of like, okay, we we want something, we want the agent to do it for us, then the agent needs to work for itself and they need to put up a bond. We actually have several teams we work with at Fractus that have agents with stake behind them trying to take actions on behalf of users and then if

they take the right actions they're incentivized in the right ways and if they take the wrong actions they lose the deposit whatever it is for example. So like I can give one example right here which is a bit more easy to understand than the second example. But the first example would be all right now the way the system works is called commitments. You write something like swap my USDC for ETH when the price is 3,000 bucks and I'll pay a 0.5% fee. You then lock up basically the assets that you want this for for this uh transaction to happen. And then essentially the agents can see that they can read that and when conditions are met they're going to try and automate the exact thing that you want at that exact time. So you don't have to be there and this is all fully on chain,

all verifiable and the guard brows are there making it deterministic. Like Miles said before then they're actually using a safety mechanism which is where I think it gets interesting. They're using UMAR like similar to like how Poly Market would use Uma for a dispute. They're using UMA for a dispute around if the agent actually fulfilled the intent of the user or not in the first place. And so if they did fulfill the intent, they get incentivized by the user with a bit of a cut on whatever that fee was they were paying. And if they didn't, they lose their deposit. But the agent needs to put up the deposit in the first place for them to be slashed. So these kinds of things around like agents working on behalf of us and having stake behind them, I think is the next era for crypto and AI. It's

still super early and to be honest, there aren't many teams actually working on this in the entire space, but for me it's kind of clear we're going in this direction. >> So is I in kind of an interesting position here? >> Yeah. Do you do you remember like um we did a whole season on restaking and part of that there was like a lot of debates in that uh in that time around uh economic security versus basically like ZK security um >> and that's going to come back right um and it's not because you can create a proof that inference ran um as it should have basically like it's uh we're zk proving like um cost I mean is it and like time is not even close there. It's

like uh and it's going to be a long time until it is. Um but you could probably prove that you know constraints outcome based constraints were satisfied right like Z if you just talked about condition triggers right uh you can prove that like the user provided these like uh condition triggers they said these are like the data sources that the data has to like be fed from that can be like proven via zkls and then you can create like some sort of you know guard rails around what can actually settle or not. Um but this is all like very very complicated and very emergent at least you know outside of Delta. Um but that

that to me I think is is going to come right back. Um same with intents and same with account abstraction. Um so like these are three like kind of uh forgotten infra narratives that were all the rage at one point over the past three or four years. Um, we've probably done podcasts on all three of them. Um, >> tons of >> And yeah, I think they're coming, it's uh, it's all coming back, which is which is pretty cool. Um, it's coming back organically, right? It's like, oh, we want this thing. We want to like this future should exist. What do we need for this future to exist? Oh, like this infrastructure that we, you know, basically forgot about for a couple years. Um but yeah that's that's I think the the thesis underpinning like IEN's

pivot and focus there. Um as well as some of the intent you know networks but >> can I point out an instance where this is already kind of taking place. It's not AI isn't related but it's getting at the same thing. So, I'm not sure how many folks are familiar with CAP the protocol, but there is basically what CAP is is if you look at the way that stable coins are managed, um you're you know, you're basically very similar to like a banking type thing for the most part, right? You have user deposits and then yield is generated in in some way.

Uh there are different models for stable coins, but um cap their innovation was their outsource. They're saying there's kind of um like the tethers of the world or the sorry the USDCs where you dump it into treasuries. Then there's kind of the decentralized the makers of the world where actually part of the way that they generate value is they allocate their user deposits um and generate yield but it's still centrally managed from kind of one team. Not entirely what you have like Spark and some of these different whatever mostly it's one team that's generating that. CAB [snorts] wants to outsource that to uh actual specialists and what they're doing is they're going to but they have to put up some collateral right to make sure that so what they're doing is they're going they're finding rest collateral from I believe sorry I don't

want to I think they started with symbiotic and are maybe now using iigen layer but see how I layer is kind of renting that out as a form of collateral management and today it's professional hedge fund managers but tomorrow right that could be an agent and I I would bet you that the margins the yield that I layer would be able to extract renting out a $12 billion balance sheet. Like, you know, let's say it's a $700 airline purchase. You can maybe extract a couple bucks on that. Um, and I bet you would never get nearly that much, right? If you were, you know, on a percentage basis if you're renting that out to a to a hedge fund, right? They're so, you know, savvy with fees and you wouldn't be able to charge very much. So, to be honest, kind of makes a lot of sense.

That's a pretty interesting uh position for a layer to be in if they can pull this pivot off. [snorts] >> So >> I agree. Yeah. Or you know I think the vault management is probably going to be impacted as well, right? Um and yeah, there was a great piece I think it was from robot VC. Um the provocative title called the Turing test was wrong. Um but I believe they're working on basically an agent operated like vault framework. Um and it speaks to some of the challenges that they incur.

>> Yeah, I think maybe just to chip in here quickly to give another example of something being worked on. I think there are many teams building super cool things right now in the crypto and AI space and like there are some more obvious examples like automation of you would be Ian layer for example giving a better UX taking a cut from that. There's one example of a team we work closely with as well which I think is quite cool enabling us to do something that we haven't been able to do before in terms of like what I could see a market developing for. So right now like I guess we have employees for example Mike and yourself uh Miles at at Delta and you know we want to incentivize our employees to get to the outcome that we want. It's similar it's like we we all

have personal intents and I think this is where it gets really interesting at the crypto and AI intersection. The question is how do you achieve that outcome? And so one of the things I do with business banking right now is you know I'll pay staff a salary like probably all of us do in here really interestingly this company that we work with right now essentially what they do is I might say hey I want to meet 10 founders and they need to be of this quality and so I want my AI to work out the founders that the intern for example was setting up calls with are meeting the bar that I want and then if it meets that bar I can basically automatically pay out with stables or something like that a way to incentivize the intern to find the founders who I actually want to

speak to and If they're not the founders I want to speak to and the calendar link has been booked and you can verify it via the link obviously then they won't receive the incentive automatically. It's like a way of incentivizing payroll and programming that with crypto as well as AI. These kinds of use cases and examples what we have today like they're only just emerging right now and I think there's a lot more value in these kinds of new flows to get to the outcomes that we want to get with our personal intent. I think these will go really really large and that will be at the intersection of crypto and AI. the things that are like automating things right now that we're already doing with bots etc. um that require less kind of crossworkflow work I think will have some value but I'm not sure exactly

longterm if it'll have enough value for us as users and in agents in future for us to really pay for for Marcos to develop around these. So also just just to finish around the layer point like I got so much respect for the layer team. I think they're amazing. The we've worked with another team also working on verifiable inference in the past and one of my learnings with infer in general especially like execution is just that it commoditizes super fast and so there are actually like many many teams working on verifiable inference in crypto. I can lay it being one of them and obviously one of the best. The question is like is that even hard to do? Like as everything gets easier to execute like we can do it on our PCs, we can do it pretty much anywhere. It's getting cheaper and cheaper. Like is

that actually long-term the way that value will acrew? Like I would maybe argue not I think the way the value will acrew is actually across you know many systems many software doing things we haven't done before incentivizing in new ways and that's where I think the most exciting intersection is right now in crypto AI. Do you think though that I mean is it about it's like charging for inference or is it renting out a balance sheet to secure uh like it almost it's almost like they become more of a um like a lending or insurance business, right? Than like a charging for than a software business. I mean that's one direction they could go. I'm speaking not knowing, right? But I mean, >> yeah, I mean, not not to plug in my boy Felix who's a venture partner with us at practice, but uh he's over at Symbiotic

and you know, I know Symbiotic right now um is, you know, kind of exploring some similar areas on the risk management kind of category and I think I layer might trend towards a similar direction. I'm sure Miles, you have some interesting thoughts there, but Symbiotic very clearly thinks that there's a market there and I would be surprised if I can layer didn't think that as well with all the stake they have. >> Yeah, without a doubt. Um, I would also uh encourage listeners um to check out kind of what I see as like the uh um outside of crypto equivalent of what we're talking about right now. It's called the AI underwriting company. Um and their approach is to basically uh be the ones to develop the sock 2 framework for agents, right? So like the basically the

canonical security like certification uh that's called the sock 2. That's what you get basically for all you know software that handles sensitive data and then they're building an insurance company off of that right so it's like if you get the you know badge of approval right from the sock tube for agents and then they're going to ensure you uh or ensure like people who are using that for really highly critical like workflows. Um I think that's a winning combo and I think that's more or less um what I and Symbiotic are trying to do. Um and I think like yeah Z I think maybe a a second maybe your your skepticism or concern and just in terms of like the horizontal trying to get a horizontal slice of like the inference

layer is very challenging. It's commoditized and like anything you do to constrain what you know consumers of those of inference want to do is is probably not a winning formula. Um and so that's that's just the challenge of it. um is like you can get these guarantees but only as fast as I can can you know support um the various models and make them somewhat deterministic. >> I would say maybe a kind of view I got wrong in terms of timing. Who knows in a longerterm time frame if if this does happen. But the verifiable inference I was excited about the last year or two in crypto. And I thought that actually the inference would be used in smart contracts and the smart contracts would go from you know more static to a bit

more dynamic and you know refreshing their state and updates a bit more over time and more creative ways via ver verifiable inference. But we haven't really seen any verifiable inference used in smart contracts at least not to the best of my knowledge on chain uh this year or 2025 I should say. And I don't know if it's just very very early, but the crypto native market for verifiable inference, which I think will come first before the other market, I think is like unclear. So yeah, it's a it might be a timing thing more than anything, but at least I thought it would be around already. >> Yeah. So maybe one example like concrete that I think will be of a verifiable inference that will be in prod and have PMF this year is AI judges for prediction markets. um you it the the

fact that we still are using like old oracle, you know, setups um for to settle prediction markets is kind of crazy to me. Um and there's going to be a prediction market upstart that realizes that they can create more markets so much more frictionlessly, right, by just having an LLM be like the collector of evidence and the ultimate judge to settle that market. um that you know introduces like another Oracle problem. Sure. Which is that you have to make sure that the LLM is running you know the model is not changing under the people under people's feet. Like if they bet on a market they want to know exactly what is going to settle it and

they don't want that to change um or be manipulated. But like that to me is a no-brainer, right? I think everything else that's like plugging I don't know verifiable inference into smart contracts is a lot you know in DeFi for example is actually like pretty challenging um or like seems like higher risk but maybe I'm wrong um we'll see I have a question for you guys on I mean the one use case that people seem very excited on agents for is some form of trading or finance and I have I have thoughts on this category myself but you know we've talked about payment And maybe not to go out on a limb, but it feels pretty obvious that it, you know, you're just going to ask your agent to,

again, I'll just use this example of book a hotel for one thing that feels less obvious is outsourcing financial decision-m or something as complicated as active trading. Not just because they could spend a bunch of money, but you know, will LLM actually be good enough there? And I I have thoughts on this, but I'm curious about what the two of you think about this category. I I'll lead quickly. I know Dave, you have more informed thoughts than I do. I think there's a very different considerations for like will uh an agent be able to manage your DeFi portfolio versus like will an agent be able to manage your multi-asset class, you know, portfolio, which is the credentiing

problem. Like I'm not going to give that agent access to my Fidelity brokerage account, nor will Fidelity like want that to happen anytime soon. But if you have a D5 portfolio, all that data is public, right? So the agent doesn't need permission to view it. Uh your wallet data is public. And then it's just about setting up like, you know, um like a trustworthy structure um to submit like an intent of what you want to do. It could be a one-time thing. It could be a longunning persistent strategy and then trust the agent enough to fund a wallet right to with your with your money to then go execute on that strategy.

>> Let's say hypothetically though the trust stuff was solved. Let's say let's say whatever that is gets solved. Will AI be managing people's money at scale like LLM? >> No, but that's one and the same thing, right? Like the tr if the trust problem is solved, then yes. If not, then no. Like that's my point. And I mean like when you say will an AI like be good enough, right? Like that is again the trust problem. Will they up? Um or can you like you know be able to instantiate guard rails that basically um guarantee no catastrophic negative outcomes. Um and until you can do that like it will probably be for like small dollars. Everybody kind of experimenting

but save Yeah, I know you lots too. Yeah, I think I agree with Mike that we should assume that we can solve for trust here. Like people are working on this, teams are working on it. It's it's hard technically, but there is some great minds at work here. And so I do think it'll be solved. Again, keep hitting this nail on the head, but it's going to be a question of how complex the infrastructure gets because yeah, it'll be solved. Trust will be solved at somewhere. For example, we'll be able to make swaps on chain. That's that's that's fine. The question is can we do things on and offchain at the same time and have a sequence of things that are being executed that we trust every single thing to be executed in the way that we expect across many different

tools and software for example that is a really hard problem. I'm not 100% sure if that will be solved and that is where the value is in my opinion. I do think that >> just swapping using agents for example isn't enough. Uh we're seeing that right now. I've been studying this a bit the last 6 to 12 months looking at these kind of LLM interfaces which I am bullish on longterm the space using it but right now it's not enough like we spoke before this podcast that all of us are crypto users here you know would you go to a chatbt like interface to make a trade like probably not because you don't trust that interface first of all and second of all like you can't really do anything extra right now so why would you go in the first place you're just going to go to unis swap use metam mask

you're familiar with it it's worked for you before that will be the case I think going forward for quite some time. Where it's going to get really interesting is like what we had with Claudebot for example this week where like okay you have Claude coowork then you have Claude bot and then all of a sudden you have access to everything in your PC locally everything across all your servers and all your contacts whatever it is and like the sky's is a limit that's when it gets really interesting when I can actually do things beyond just make the trade what I can already do in crypto I don't think is enough right now >> right and I think that is the challenge right there's uh maybe like the two parallels are for shopping something as that seems as simple does that it's uh limited by merchants that will accept

you know stable coins or accept um agent payments over legacy rails and hopefully it's the stable coin side. Um, and for finance it's like, you know, access to asset classes, right, that the agent has. Um, and as long as it's just, you know, limited to like open systems for DeFi or like open systems for payments, you know, it's going to be kind of slow, right? Um, >> well, I have I have like one I mean maybe to push back a little bit on I think there's a difference in like Miles you said that the trust problem and will people use agents to manage money are one and the same. Like I would actually push I'm not 100% sure I agree with that because like let's use the example of a

trader at a hedge fund. Um let's say they underperform there there are very different ways with very different career repercussions that you can lose money. One thing is you like fat finger a trade or you blow out a risk parameter. You like you you know you go balls long, you know, 20x leverage on Tesla. Um like you you kind of have these breaks in procedure and you lose money that way versus maybe you just suck as a trader and you can't consistently outperform and you end up slowly getting fired. Those are the two different ways that LLMs can like the former category that I just described is like kind of the trust and infrastructure problems, right? your LLM hallucinates and like 100x over your spend limit or something, right? That's

one way that they but there's also an there's also a possibility that LLMs actually at scale can't do what like a brilliant long short hedge fund manager at a New York firm can do. And actually where what I would describe as the limitation there is the amount of data that is online. um like LLMs are ultimately limited by the amount of data that they have and it's possible that they might just never be granted access to the type of data that would lead to persistent alpha and but one you know where one category that I'm super bullish on for what agents are going to end up doing in a market they're going to eat is raas I see this as like >> already happening already happening for

sure >> like betterment and those sorts of I mean just no offense any RAS that are listening, but I think it's kind of the last vestage of how finance used to work, which is like I want to be able to look someone in the eye and shake their hand and know they're managing my money, you know, give them 80 pips a year for that honor. It's like that is that is toast. That is and not only because I mean mostly again I mean I would just never think to use a wealth manager advisor like I just that feels like such an antiquated concept to me already. Um, and AI can already, one of the advantages that it has is it understands you really, really, really well. And my kind of understanding of

RAAS is it's not always just like the optimal positioning. It's like a sales job. It's like, oh, you like biotechs, like here, I know what you like. And they like it's very customizing for you, right? You want to feel understood as a as a consumer. Like, oh, I don't like stocks. I'm in all bonds, you know, when I'm 25. Okay, it doesn't make sense, but here you go. You know, I think that LLMs will do a better job than any individual salesman or, you know, RA portfolio constructor could do because they'll make you feel heard and understood better. And there's not really any alpha in that business anyway. So again, sorry any RAS that I've offended. I understand RAS is a broad brush. are institutionally focused ones with managed accounts and all that stuff. But like the broad retail basically passive money index construction that is just

going to get like that that's going to get eaten in five years. I would guess. >> Yeah. I think Mike the word you're touching on there is they offer a personalized experience. And I think any human >> any human industry right now which offers a very personalized experience is bound to be disrupted by AI because that's where AI really excels and shines. Um I I personally think um that you know the largest markets in the crypto and AI intersection won't be driven by novelty. I just think they'll come as a relief to users who find like current systems cumbersome. And so what I mean by this is I think the bigger pains that we have today, we still have so many pains when we use all this stuff including crypto. By the way, crypto is

a massive pain to use. You know, one use case I have which I've been talking about the last few weeks is like I have you know some treasury. I want to put it on chain. I want to diversify the strategy. Like why is it so hard for me to understand like what the hell is even out there in the first place? What I can execute on? Is there enough liquidity for an institution? Etc. etc. Like super. And then I can't execute it with one button. I have to then click everything myself. there are transaction fees. Like this is really really annoying and this actually kind of pisses me off and so like I'm willing to pay for a product that solves this problem for me and I imagine there are probably others that would do the same and there are many teams working in similar directions but like no one's

perfected the understanding of the user itself like what I Zave want and then alongside everyone else then try and make a product based on that like I'm sure you hit the nail on the head with that. You know what? You said something that actually made me laugh at the beginning of this save, which is you're like the big difference between the legacy system, it was designed for humans. So, what that implies is that crypto was not designed for humans and it absolutely [laughter] was not because it's so difficult to use and such a pain in the ass. But this is, by the way, we didn't really describe what like X42 and some of these standards are actually helping with. And it's exactly that like if you go down the pain points of like imagine what a what a standard has to help an an agent solve. It's like

everything that you, anyone that's ever done anything on chain already knows what the pain points are. You have to pay gas fee, right? Like depending on where you are, like sometimes you have to make an estimate. Um, you have to manage essentially currency risk. You have to bridge, right? Finality is different across different chains if you're ever bridging. I mean, there's a gajillion, you know, every painoint that you have dealt with, someone's gonna have to program a protocol in a standard for to direct agents through. Um, so yeah, I I think that this stuff is going to end up getting built 100%. And yeah, I mean, remember, you know, how we've been calling for crypto needs to improve its UX forever. It's probably never going to improve, but it's just not going to be you that needs to do it.

It's going to be agents. And there's always the intense too, right? >> There's also a timing thing here, I think, because you mentioned like X42 is built for pains that we have today. I have to admit, like I'm not sure if I have a pain right now of like needing to pay microp payments on chain. Like I think that's actually kind of a future pain that might come about from agents needing to do this by calling APIs for humans using these models for example. Like that makes a lot of sense but like as a user right now there's actually not enough third parties on the other side almost offering X42 standards for my agents to tap into. Like it's it's still being built out and I think there actually may be future problems but I think there's a lot of current problems that we have today that like some things

can still be built around that make so much sense in the crypto space. I always had a kind of view that usability in crypto was, you know, terrible. And I think we all agree on that. And I think AI can help us with usability and discovery and engagement is what I used to say back in the day around like personalization. Like a lot of the applications we use in crypto, there's just some obvious places you can use personalization. Like we had a product launch that we work closely with called Fireplace yesterday and like it's a prediction market app. But imagine like I'm making money on markets and then the AI is understanding what markets I'm making money on and it keeps recommending me markets and I keep putting on trades there and I'm making more and more money and it's helping me trade even better etc etc. It's like

very personalized experience. It's like raia for prediction markets if you want to put it like that. So I think there's still a lot of opportunities at the personalization layer as well and crypto application layer. >> Yeah. I'm curious like maybe to close here um if you have any big conclusions about how the world changes in the world of even aentic payments and commerce. And the thing that I always come back to is how is this going to impact KYC? Because a huge advantage of crypto is that you don't need to do something like KYC and you know I'm just looking at some of these broader macro trends and like everything's global from day one. I mean the the way that I at least like to describe payments and commerce on uh crypto versus in you know agented

commerce and crypto versus non is in the legacy payment assumes the legacy payment system assumes that you are operating within the bounds of legality and there's legal recourse and everything is cryp in crypto is built with the opposite assumption and so you kind of have this huge decrease in friction and globality from day one and I guess the big question here really is like KYC Like that's because for the f there hasn't really been a super strong market demand to overturn KYC and if there were to be adoption from agents with all the AI money and hype and competition behind it and there was a huge market force business case to not

have KYC like honestly I've never really been bearish KYC because I don't see what would reverse it. This is kind of the very first thing that it's like, huh, there could actually be a legitimate business case to be made, which is how we know that's how the US works, right? You make a business case. Um, and yeah, that I guess that's one of my big conclusions from understanding a little bit about the frictions and how you know if this system were to get adopted. >> I mean, I think I think like you're kind of touching on like the key bottleneck, right? Like execution is now incredibly cheap. you can build anything and like these AI agents like 99% of the time are gonna like do what you want them to do even if it's with money assuming they

have the ability to spend it you know um but it's like everything that happens around execution right um and so these are like the policy checks that you will like you know get in a lot of trouble for if you violate um it is like the you know we were talking about like the hedge fund trader like earlier right like they have obligations to their investors to be doing the right thing all the time and that it everything they do goes through a bunch of sets of like human approvals right and so I think like >> the as things get automated and as things like are shifted over to agents I think institutions will actually demand stronger guarantees than they have now because they are going to be pulling

humans out of the loop and if they don't pull the humans out of the loop we're not actually getting to this future right we're not actually like um I don't know breaking through like the the uh getting like the real gains that we could see, right? And so I think there we'll need to like automate verification of all of those things that happen around execution um that humans typically have to do or like typically a human in the loop and I think that's the the hard part. Um, >> yeah, I I have a call to make here which I also mentioned pre-podcast, but going back to your KYC uh question, Mike, uh, you know, we have crypto. One of the awesome things about it is that we don't

need to KYC every time we use an application and that's fantastic UX. Um, really it's getting past wool gardens in finance. I think in AI, W gardens also exist and they're going to get worse over time. So like OpenAI, you know, Anthropic, everyone, they're going to continue trying to build these closed gardens and make it really, really hard for people to access anything outside of those wood gardens. And so I actually wonder if a lot of the value at the crypto and AI intersection will be getting past wood gardens of AI itself. And to give an example here, again going back to Claudebot, which came out this week, there were some stories and reports of different AI companies shutting down personal accounts just for accessing things via Claudebot because they want to direct the traffic more

towards Claude, for example, or CHBT, etc. And so if Quadbot delivers us a lot of value and we want to access these services, but we're being blocked a bit like finance used to be back in the day and still is to some extent, well, there's going to be workarounds made, especially in the crypto and AI space, which I think will be incredibly valuable to give people what they actually want. And that's where I see a massive opportunity, that intersection, which I'm potentially the most excited about right now. >> I agree, man. Future is bright. Future is bright. All right, guys. Um, I think we can probably end it here. Uh, we're just under an hour. Uh, but this has been a ton of fun and hopefully this has sparked some interesting conversation and yeah guys, if we missed anything on the crypto and AI front, um, join the

bell curve chat and drop and let us know. But this has been a ton of fun, guys. Miles, Dave, appreciate you both. Cheers and see you next week.