Why AI Agents Need Guardrails: Ade Adepoju on Latch and Agentic Stablecoin Payments

Alright. Alright. Welcome to the show.

Hey, guys. How are y'all doing? Really good. Really good. Happy to have you. Likewise. Likewise. Thank you. Glad to be here. We're our our little intro video there is talking about the builders who are already climbing towards Adjantic Finance, and you already got a builder. I'm like, I need to start drinking more after watching that. But It's mean, this entire work from home, work from home, you know, WFH. So I can That

sounds good. Yeah. I've been you you have been trying to think through, like, where is there a possibility to film the show in the wilderness somehow, like, with Wi Fi and what that would look like? So we'll let you know if we figure that out. Honestly and and Drew knows this, but I'm always scheming up, like, how do we get some of these builders together? Like, vents and stuff. Maybe we'll do one, like, in Jackson Hole or something in the mountains. We'll we'll go for a hike. I would come. I would totally come.

I think right now, you know, we say the bear market is good for building, but it's also a good time to reaffirm,

you know, the relationships that we've built. So I'd totally come. Just let me know when. Let's go. Sounds good. I love it. So bring us bring us in a little bit to to what you're building and and how it relates to this this conversation around agentic commerce? Absolutely. Absolutely. You know, it's a there there there's a bit of a curse that we have internally where we tend to be quite early to things, and then we spend so long trying to explain it, and then it turns out, you know, now is a good time.

So when we just talk about the robot blockchain, we you know, my my background came from distributed systems, building for humans, and then I went to Mistin where I was one of their earliest employees there and built out that blockchain. So we phenomenal team, phenomenal vision. But what I noticed was there's a bit of a problem around the integrations needed to do anything in crypto. Talk to anybody right now.

They need to spin up a team that understands the bitterness of smart contracts, then sit and negotiate with Chainlink, figure out how to work with LayerZero. I was like, there's all these things that are cool, but they actually slow down adoption. And they so you burn a lot of cost. Rialo's infrastructure was like, what if you just said blockchain there was no Web two versus Web three. Like, Web three likes Web two. Okay? You don't need to start from scratch.

You have smart contracts that can call APIs that understand all the stuff we built in the financial industry. Like, there was a wave of digitization that happened before crypto around crypto that everyone did. And then when they come to crypto, they have to be another way of digitization. That switching cost is too high. So Riello's core thesis is simplify the stack, delete all the unnecessary components. It's not about TPS.

It's about speed of integration, and make APIs a primitive, which means smart contracts that can directly talk to plant can directly trigger stuff in Web two point can trigger a payment. Now you don't need to make those guys understand crypto. Crypto understands that. Right? It's it's flip it around. Make the blockchain itself world aware so it integrates with everything that exists. This is good for agents because now your smart contract can directly trigger talk to earnings directly from the smart contract.

It's good for stablecoins. It's good for pretty much anything that has an API, which is everything. Right? And that was kind of our core thesis. And now it's kind of a great time because we're seeing a lot of interest, especially around something we built called Latch. But, yeah, I can get into the details of But the core premise is don't come to crypto. Let crypto come to you. And once you get that, things get simplified. That's interesting.

And it seems like something you were seeing at MistIn is this sort of problem of too many solutions.

Like, you're it it sounds ironic. Right? But, like, there's so much coordination and switching costs because it's like, there are 10 solutions to solve the one problem, and this is about sort of a simplification and a unification. That's your core Yeah.

Yes. Yes. So, you know, again, this is not a not a problem or this is a every new prop every new technology goes through this phase. You get an explosion of things. I'm sure you guys remember, I don't if you're old enough to remember the the ports on the back of computers. There's, like, 50 different types. One for a printer and one for your toothbrush and one for whatever else. And then you get a consolidation like, okay. USB. Right? Same with phones. All kinds of things in it. So it's normal.

Even with AI, there's all kinds of models and stuff. So this is just the natural apps do have on your phone? You know? Exactly. And how many do you actually use? Right? So it's like, you get this natural explosion and consolidation once you figure out what works. We saw that consolidation was inevitable in infrastructure. There's gonna be a shrink at an 8020, and that's kind of where we were positioning. The problem with crypto is the existence stuff. It's hard to just evolve them.

The economic economics don't make sense. Right? Once you've launched a chain or an infrastructure project, it's very hard to turn it around. You kinda have to start with that goal in mind of integration. The incentives don't align. You could imagine if Ethereum decided to integrate Oracles, right, which they thought about. Now they're not gonna be in a good position with maybe like a chain link, for example, theoretically.

My point is, like, you already have these entrenched relationships that are hard to revert theoretically. So we wanted to come in from a different angle of assume the world is about consolidation and integration of infrastructure. What would that world look like with that goal in mind. That's kind of our angle. And we see it happening already. Right? Yeah. In many places. Yeah. So two two part question from my end. Like, what, yeah, what is happening from a consolidation standpoint?

Like, what is getting consolidated?

And are there learnings that you are looking back? I I I love the App Store, you know, at least in my mind, that kind of helps me think about it, where you have all these different apps that were, like, doing one thing really well or ports on the computer that were doing one thing really well. Like, your audio, for example, like, now we have AirPods. Like, we don't even need that anymore. It's just unnecessary confusion. What are you what are you seeing thematically that typically gets consolidated?

And maybe it sounds like that's already starting to happen in from what you're seeing. Well, the first things first, we definitely

we're seeing things like you know, we when we announced this a year and a half ago, people said we were crazy, which was you can sort of enshrine a lot of components, like oracles to some extent, automation to some extent, and then megathons, you know, enshrine oracles some extent. Tempo announced recurring payments. A lot of these features should just be table stakes. But there was an overly purest mentality that a blockchain is only a ledger. Well, customer doesn't care about your philosophy.

The customer cares about simplification. Right? But this actually enables more, and we took it to the extreme. This has enabled us to build something quite interesting called Latch. K? And I'll talk to you about Latch. Latch is powered by Realo. What it does is it uses blockchain properties to give you guardrails around agents for payments, for actions, for spend, etcetera. This is the extreme case that we're excited about. What does this mean?

This means we've spent about a decade, over a decade now saying smart contracts, you know, they provide rules and they're bulletproof, etcetera, etcetera, etcetera. But we also know that in the real world, we have a problem of agents going rogue, and we wanted to give them, like, rules and policies. What if we took all the stuff we've done for making blockchain secure and trustless and fair and protected to apply that to agentic actions. Exactly. And that's Latch.

So Latch allows us to use our core infrastructure around agentic access control and permissioning, which everybody's talking about nowadays. All the the GitHub exploits was a credential leak problem. Employees are dealing with this every day. Every employ every team has at least I think the ratio was one to 45 number of employees, human agents, humans to agents. But you don't actually know what's going on behind the scenes. You don't know how to give them access. You don't know how to revoke access.

You don't know how to track costs. So Lotte was a is a product that allows us we needed this internally, and we're actually opening this up to the world. We have a waitlist of, I think, about 23,000 people right now, and people are already testing it out and some interesting partnerships on the way. It allows us to use our our learnings and infrastructure and blockchains to provide guardrails and also then enable agents to do more. Sorry for rambling, but, yeah, that's the No. You're good.

This is this is this is the privilege of being a guest. You can ramble. So how does it work? We can actually walk through these four things if if that's a good idea. We can kinda go through these four. And and we have a problem at Sub Zero since day one, which is our docs are always way, like, a year behind our progress.

Some of the stuff is team is shipping real real.

Yeah. Some of the stuff here is already old. Happy to do a walk through personally separately, but here's the core idea. We modeled the Internet off of a very bad primitive, which was fees. Right? Like, I give you a key, physical metal key to my house, and you're the plumber. You're not supposed to go to to my bedroom. You're only supposed to go to the bathroom to fix a pipe. That's it. But the way keys are, keys typically do not have rules around them.

What I would like to give you is a key that only lets you go to the bathroom for the duration of this job you need to do as a plumber. And once you're done, your key is gone. It's destroyed. But keys and access today are way too permissive. And API keys could if I have access to your crypto wallet key, I can do whatever. What we wanted to do was can we do customizable, flexible, and controllable access policies? So you never have to give away the master key. You never have to give away the raw credential.

I give you a scope permission that only lets you do exactly what you're supposed to do. That's the way the world should go. Now when it was for Schulman's, this was fine because you have the law. I can sue you. You know, you don't do too many crazy things. Agents can spin up and may do damage in seconds before you can react. So we noticed this internally that there's all this sprawl of access. So Rialo uses Latch uses the Rialo engine to define custom flexible rules and guardrails and enforces them.

So it basically sits in between and goes, hey, your agent wants to do x y z. Is this fits your policy? And if it doesn't, it goes, you're not allowed to do that. So and you can you don't even have to trust us. Why? Trust the smart contract. We've that's the thing that we we know how to do very well. So it's basically a system that sits between agentic workflows and real systems and allows you get governance. Now here's where we took it a step further. We do our checks at the hardware level. Why is this unique?

Software is getting poisoned every day. So even though you may see some implementations to do it in software, you cannot always get guarantees that the software is doing what it says it's supposed to do. With our approach, the hardware itself, which is extremely difficult to break, is giving you the thumbs up that this agent's action is approved or disapproved. There are solutions of people that have tried this in pure software, and the backdoors mean that it ends up not working. Right?

So our approach is probably the only secure one we know of, and we're even experimenting with some really elaborate solutions. And there are some names that we will announce over the next few weeks that are codesigning this with us, but we believe that this is the next way in which agents that handle real money will have to be governed. Otherwise, it just become dangerous. Pulling on that thread a little bit, Ade, the Yes. Software can become poisoned. You see that hardware is a little harder too.

Is this an enterprise buying your hardware? How how is that working? No. No. Quick question. So it's something called the root of trust, which is what is the thing that giving you guarantees? What is the thing that making you you know, that giving you the a guarantee that something is doing what it's supposed to do? As you can see in this diagram, there's an OpenAI key, and there's an Angel Ventures, an engineering bot. There's all these helper agents.

How do you know that this engineering agent, this engineering bot actually did what you said it can do? Like, is limited. They've if if you trust in the software, the software itself, and you can check with them like, so Team PCP, that's their name. There's rogue hackers out there poisoning software in a way that you would never know. Hardware, on the other hand, is called an attestation. It gives you a fingerprint. So when you run through latch, the hardware will give you a fingerprint that's unforgeable. Right?

So that means our latch is hardware backed. So you're not buying our hardware. It's still a software exactly. Right? It's still software, but at least the root of trust, the thing you're trusting is hardware built by well defined hardware manufacturers that are guarding guarding other stuff. It's not just real little sub zero software. Right? So you can verify that yourself. This is where we believe the world has to go to because software is moving quicker than we can react.

Hardware and the Internet seems this seems extremely concerning. Like, GitHub It is concerning. Typically. I mean, that, like, how git.

It's like the worst person that I would wanna hear is getting victimized

by this group. If GitHub that's in charge of everyone's code is not secure, then you should really be concerned about what's going on. Right? So, yeah, it's really concerning. Yeah. Laughing and chuckling is, like, the only thing you can kinda do because Laughing to hide real pain. Yeah. Dude, it's it's it's crazy because, you know, we we we do this for ourselves internally and realized there was nothing out there. But, thankfully, we have knowledge on formal verification, smart contract, hardware.

We're like, what if we bring this together and solve it not just for us, but for others. Right? So but to be frank with you, the thing is things like this are hard to explain until it's you. Yes. That's that's that's getting problem problem. Right? So we're trying to make sure people understand that at the velocity at which things are going, it's not just to protect yourself. It's also easy to fall out of compliance. Compliance today,

you know, you just talk to it every year. You reverify that you haven't done anything bad. Within two mic milliseconds, an agent can violate every compliance rule at the company. So we're excited about this world where Latch can then enable continuous compliance. It's always checking and making sure your systems are up to date. You're doing the right thing. All your APIs are ready, are doing everything they're supposed to do.

Are you when you it sounds like you have some of these, like, to be being candid and transparent. I just got a text from my mom. Hey. We're stopping by the office. Dude, can we come in? I'm like, mom, I'm literally in the middle of a live show right now. Come on, mom. My mom always picks the best time. Mom yeah. Mom always knows when to pop in. Question question is around the actually, the usage of this. It sounds like you've kind of gotten some of these early beta users. I'm looking at the site here.

It looks like you guys are marketing this as a beta at least right now. Without naming names, it sounds like there's some things you're excited to to announce, which we'd love to help amplify some of this. Let us Yeah. What are the use cases? Like, who who who are you excited to use this most?

There's personal, and then there's business. Right? I mean, my personal view on things was always, like, five years ahead. So there's a few things that are quite interesting. And, again, interesting, not necessarily business. So I'll talk about the interesting part first. Do you know that APIs have terms of service, but you actually don't know what the terms of service are? You have to read another document to know what the terms of service are. Agents now need to understand multiple things to know what they can do.

With Riello and Latch, you can encode terms of service rules around how to use systems directly in the API itself. Like, you don't need a front end. So I'm excited about a world in which machines can read rules, legal documents, conform to them, pay for them all in one place without having to call somebody else. That's the more futuristic one, but it's quite exciting. But the agentic use cases are phenomenal. One of the things we haven't pushed out yet is called machine machine based identity.

We invented this thing that lets you not approve an agent or approve a person, but approve a machine. Why is this unique? So let's say in your company, we're seeing this used, a lot of people want this. Instead of ever having to give someone a password or an API key, you've already given them a laptop. Why can't you just say these people with these laptops are the only ones that can access some services?

This means if they ever leak a key, it doesn't matter because the key is only gonna work if it comes from their laptop. So this GitHub issue could have been mitigated with machine bound identity because even if you put all your keys in GitHub, someone will need to physically get to your house, get to your machine for the key to be useful. So this is an interesting thing that would sign in. It also helps you civil and agentic commerce. Right now, people use wallets as a way to identify agents.

That's not a great idea. An agent can set up a million wallets. You won't know which is which. It can pass around the keys. But what an agent cannot do is go to the Apple store and buy a new laptop and spin that up. So now with machine based identity, we can use the actual hardware itself as a way to prove what an agent is and where it's running from. That's much harder to clone. It's very hard to clone a MacBook Pro, but it's very easy to create a new API a new crypto key.

So we're seeing people wanting to use this. And then the third one is continuous compliance. We think this is gonna be quite revolutionary. Latch allows you to it can basically listen to your Slack and everything you're doing, and it's constantly making sure no matter what's changing the energetic world, you do not fall out of compliance. This is good for need for for regulated businesses where, like, whenever we do business with some folks, they have to check that we have checked all the boxes.

We're not leaking information. You could think health care, HIPAA. There's all these things. So those guys are not adopting agents because agents can make them fall out of compliance too easily. So you have this chicken and egg problem. They want to use agents. They wanna use AI, but they have no way to control the AI in a way that's safe. We can enable them with Latch. So we can tell them, hey. You don't wanna get all this cool fable and all these cool AI agents. This system is not just about protecting you.

It's about letting you do more. So in those industries, we can do quite a bit. Now compliance, security, all this stuff is not inherently sexy to some people. So we actually found unique ways to make all this make sense. So we're going to be announcing something soon that's a gamified way to let you play around with compliance. So it's basically a game. It's modeled after I I I raced cars as a hobby. It's modeled after car racing where you have restrictions, but you have to beat the game given restrictions.

No no token maxing. It's not about just spending infinitely. It's about understanding that nothing's free and trying to achieve tasks given policies and guardrails. So this is something you'll be seeing a lot more of very soon. We also use Latch to do our internal token send management. Because of a policy engine, can check what someone's about to do. It can say, hey. You don't need to use OPUS 4.8 to ask what the time is, and it it reroutes what the people do.

So we're also seeing it for agentic payments, x four zero two, MPP, etcetera, where people wanna issue flexible spending cards. I'm not sure if you guys ever know what's in a company that has a spend expense policies. Those things are a nightmare. It's a 10 page doc that you never know if you're violating. Like, I don't know if I can go to the Ritz, but it's like, you know, it's very complicated. The problem that, like, the ramp and the Brex layer solves. Right? Typically? To some extent.

But what does it look like for agents? Remember, these things do it after the fact. You spend, and then you make do your expense report. We said move all that into the code layer. So before the agent even spends exactly. Before the agent even spends, the latch runs and says, you're not allowed to do this. Like, you're you're offered. So we've been able to move all of this into the agentic card. So you can imagine a policy based card that you issue to employees,

agents, etcetera, etcetera. So we're seeing a lot there. Yeah. Yeah. And picture when you were talking about that one to 45 ratio, I was imagining onboarding, like, 45 college interns and handing them all cards.

That that question. You know you know You're good. I just changed the layout. Oh, I was gonna say I changed. I was like, wait. Where did they go? Sorry. I'm not I'll bring you back. I'll bring you back. No. No. No. It's okay. I was like, you're not any you're not friends. Knowledge on us. All my hiking buddies are gone. Okay. Anyways okay. Okay. So so so, basically, we we've seen this a lot even internally, and I should not admit this on Camden, but it's true.

We, you know, had an employee that left a while back, and they had some API key, but we were also had the API key. Was, like, five people using one API key. Right? Because it was a cloud key. But they happened to have it on their machine completely innocuously. And part of their they were still kind of using it because they forgot. And you don't know. With agents, it's gonna happen where everybody's sharing a resource and you don't know who has access to what. They leak stuff. So this is inevitable.

And the other part also is with some of this stuff, we are making call it the bad trade. The trade is, you know, you're seeing companies switch out employees because their token spend is high, so they need to offset cost. But it's a it's like you're replacing employees that you can manage and understand with something you cannot understand, you can't manage. So you don't even know if if this is a fair assessment.

So we think the world needs to have controls and guardrails for agents, and we are proud to be supporting this. Over the next few weeks, we plan to announce quite a lot. We'd love for you guys to participate in the use cases. And We would love to also get you in touch if you haven't. Have you met the Arculus team by chance? Have you heard of what we've doing? No. The Arculus team is really interesting,

and we've been working with them for a long, long time as we've built the business. And they their background is in, like, Adam Lowe, the the CEO of this business or the head of product innovation of this business, really is focusing on similarly, like, the the hardware side of signatures. And they actually produce and they're wholly owned by a company called Composecure that produces, like, 95% of all the metal cards, like like credit cards or debit cards in the world.

And they've built this really interesting technology around signing, and their backgrounds in, like, military and defense and applications for, like, physical signatures of access control and all these other things. They're bringing that level of that they're really trying to figure out that that technology and how it applies to signatures for payments and other financial use cases. So given what you're building, I think that could be a really interesting connection.

I I I wish I can disclose too, but I just had a meeting before this that I I really wanna talk about, but You've already been disclosing so many things. Why why not this one? You know, I thank the guys. You know, I don't want them to sue me. I don't know. But but but objectively

Their agents will sue me. Their agents will sue me. It's it's inevitable. You know why? Here's why it's inevitable. A lot of yes. Very true. So we we were talking to them about this interesting angle of a lot of the things we've done today, like the fraud detection, all this stuff, is actually because of lack of information that does not propagate. So when I make a payment on my laptop, my laptop knows where I am. It knows all this information, but it never makes it to the other end.

So you have all these fraud detection mechanisms. When you have hardware based signing, hardware based pool from payments, the merchant can rest easy knowing that something of trust originated this. So this is inevitable also for agents. You need the roots of trust to be something that has a high bar. Right? But I'm curious. What are you guys excited about nowadays? What are you seeing? I'm sure you you you you see a lot of, like, crazy people like me come out and talk about the future.

What what's on your minds, Zach? What you got? I mean, this is one of the categories is taking what I I I guess I just think back to the adoption of ChatGPT, for instance, and I think about how it was like it felt like a snap. And I think sometimes great products or great innovations, we have this idea. It's like, oh, there's really slow adoption curves. And, of course, there's sort of, like, this long buildup to making that happen.

But I think the skeptic in me when I've heard agentic payments, agentic commerce has thought of all the blockers and all the reasons why not. But then thinking adoption of how fast it went from this chatty bitty thing is weird or like it won't catch on or, you know, all the models. Right? To you're telling it your deepest, darkest secrets and like asking for advice on your personal relationships and whatever

else, you know? Now Sam knows what what you think about it.

Exactly. Sam Altman knows about us all, and enterprises are adopting and all this kind of thing. And so I guess it's on my mind in this sort of agentic economy category of how we might go from, oh, yeah, Ade, kind of this crazy builder guy to, like, nope. He was the prophet saying the right thing all along. So that's very that's very top of mind. Well, you might be a prophet. You can tell us when we're hiking. Okay. It's okay. We'll I'll the account of Ade. Yes.

Well, we do need to wrap here, but, Ade, it was great to great to have you on and meet you. Just love your energy and the the expertise that you share with what you're building. So don't be a don't be a stranger. We're this is actually the first day that we're launching this specific segment powered by Breeze, and it's really focused on this conversation around the moment around agents and money and how those things interact together, and that's something that we're gonna continue to cover deeply.

So whether it's you or other people in your partner ecosystem or client ecosystem doing cool stuff, let Zach or myself know, and and we'd love to keep following following and learning on our trail. I'll do that. I already talked to you about the hardware I talked to. I'm trying to talk to them about that. Yeah. Arculus. Go check them out. Happy to intro