Enforceable Guardrails for Agentic Finance

Transcript

[applause] >> So, I'm glad that Cypherpunks are dead. I was one of the original Cypherpunks and I'm, you know, was a little worried there for a moment. Um as she said, I'm from Dean Tribble from Agoric and we're building a product called Wyre Max, but I'm really going to talk about the Here we go. Agents. And what does it take to do agents? Um so, if people have not been paying attention, um financial activity is moving dramatically from people engaged in all of the loops to agents being able to execute on behalf of people. Now, there are major obstacles to that and a

lot of the talk is about how to solve those, but we know that agents will deploy capital. And and there's not just like Giza that is an AI that is deploying its own capital, but um Fireblocks, Coinbase, Circle, uh this morning MetaMask, all rolling out APIs to enable agents to be able to talk to those systems. And this is not, you know, and this is largely driven by the desire for smart automation that can adapt to the complexity of our evolving economy um and and operate at the speed of that electronic economy.

And this is not reversible. I mean, it is just in its infancy. There's There There are relatively small amounts of tens or maybe a few hundred million dedicated to this right now, but it is a nascent market that people view as inevitable. This is why everyone's getting on board is because of the advantages and the financial world is already too complex to manage by hand and it's just getting more and more advanced. And so, agents are how, you know, you and I that are not professionals, you know, at directly managing money, get to keep up.

And the market agrees, right? I mean, it's it's they're wiring agents into uh not just advisory capacities, but into execution capacity. So, that means that agents will uh will be transacting directly on the money. But, there are key obstacles to this. You could see the rise over the last several months of issues around agents going off the rails, whether it's deleting production databases of running services, or it is uh there was a agent that no infrastructure problem, it's just that the agent got talked out of its money. And because it had the

authority to send its money to arbitrary destinations, the attacker could persuade it to send its money to them so they could help. And uh many of us have seen those sorts of phishing attacks on us. Agents are just as subject to it. And so, we know, and the industry knows, and the market knows that you can't actually trust agents with your money right now. That it's an And that is the number one obstacle to a world where agents can operate on our behalf 24/7 to let us cope and grow with an ever richer economy.

And so, what can we do to make agents trustworthy? Because that is the number one problem to solve going forward to grow into that economy. So, there's a lot of things you can't do. There's a lot of things that don't solve that problem. Right? The the in much of financial industry, you allow people to execute, and then you audit them after the fact. At the speed of computers, at the speed of light that our that our economy is moving towards, that's too late, right? The money is shipped. It has been converted, it's been traded, and it's off to some country that we didn't want to be building missiles with it, right?

And so, monitoring is too late. It catches the It catches the theft, the fraud, the simple mistakes after it's no longer possible to recover from them. The other thing that you see is uh is that, you know, people give actors the rules, right? They tell the actor, "Okay, you know, you can only put your money into these vaults. You can only put your money onto these chains. You can only send your money back to me." And then, in order to be helpful, someone persuades persuades it that it's much cheaper, better, faster to send your money back to me by giving it to them first. And then, somehow I never see the money, right? So, so trying to get the actor that we want to rope in and and and the agent to act correctly,

and just say, "Well, that's how we're going to make it happen." That doesn't work. That's the thing we're trying to police, right? And that's like giving a junior clerk that you just hired the discretion to decide when they have traded enough or when they have got enough leverage. And there are many banks that have failed that can tell you that that's not necessarily a good idea. And the other thing that you see, and we saw this at a presentation from an exchange talking about how you can get API access to trading on that exchange, is they have additional agents. They either have strong guidance to your current agent, which is that second item, or they have additional agents that will look at what you do and decide whether it's approved.

And those agents are no more reliable than the agent that is issuing the transaction, right? That's not the kind of deterministic, high-integrity that you want to be protecting access and execution on real funds. And so, you basically more of the same won't help. So, there needs to be a change in the structural environment to make that happen. And you've again seen this increasing awareness of the guardrails problem as something that everyone is now leading with when they talk about agents, but many of them talk about it as well. I said, I've got this smarter agent and it'll it'll do better, I'm sure.

So, the answer is not more of those. It is direct enforcement of a mandate of you know this financial industry, this idea of bounded authority that this thing can execute in no matter what it tries to do, right? It's about what it can do, not what it wants to do. And so, that is not and and you don't achieve this by giving it keys and hoping that it will do the right thing. Right? That's like giving signing authority over your bank accounts to the person who's off shopping groceries for you.

Right? You wouldn't do that. It doesn't make any sense. You want them to have confined execution that no matter what they do, they can only spend so much money. And really you'd like it that they can only spend it on the kinds of foods that you asked to buy, right? It's also similar like Robinhood released an API for doing trading where you carve off some money and give it keys to that. Well, giving it just less money with keys is like carving off an account for your kid. It's a great learning experience. We'll learn a lot from that. But that's not what you're going to do with your retirement account. Right?

That's and it's not what you want to do even inside of a company for allocating funds. That's just a place to learn. Very powerful, very glad they're doing it. It validates how important this is as an industry motion, but it isn't at all a solution. The second way people address this is custody. Right? So, that's basically I put my I I give my money into Coinbase Anchorage Fireblocks whatever, and they will provide the execution. I can do rules in that context, and then my agent talks to their API, and it can only do the things that I set up in rules. And that's really good. And you know, and and banks are starting to add that. Mercury added something like that, and so forth.

That's really useful and valuable, but custody providers are implementing non-custodial solutions because custody cannot move at the speed uh that that of the economy. It cannot move at the speed of computers. It cannot move at the speed of light, right? And so so again, that's valuable and an important niche, but the future of agentic commerce, the rich, dynamic, adaptive economy that smart contracts, crypto, and even uh um uh traditional trad-fi services or trad-fi services will provide, custody is not a solution. Or it's not a complete solution.

And again, you can't do it with terms of service, right? You can't do it with, you know, I'm going to I'm going to sign and and and and and sign off on the risks and sign off on the rights, and my agent will follow those rules. Again, that's just doubling down on the thing that we're trying to prevent. So, how do you enforce a mandate? Okay? Well, we've got this tool, smart contracts, that solves this enormous what's called the principal-agent problem, if you're familiar with this in in uh economics literature. So, what do we want? The user, possibly talking to an AI, decides what the mandate is. Decides what they want to do. I was talking to um some

folks from Morpho yesterday, and a common one they get is, "I want my money sitting in this, you know, this relatively low-risk, you know, 6% vault until some vault goes more, you know, it up to 8% or more. And on a regular basis on different chains, vaults will spike up to 15 or 20% because there was some big liquidity demand, right? The The demand for USDC liquidity, USDT liquidity changes which chain has the highest demand as the industry evolves. And so the the the mandate I want is put it in this particular vault with a reasonable flat interest rate until that spikes up to 8%. Then move my money there for a few

days and then move it back. That's what I want my strategy to be. And the mandate part of that is here's the set of vaults you're allowed to participate in, and here's the maximum you can put in any one vault. And then the strategy of how quickly to respond and and and how quickly to pull back out and whether the motion is enough and is real and let me look at the Twitter signals to see whether the reason why the interest rate spiked is because there's demand or is it because they've got a liquidity crisis. Um and based on all of that is what the agent will do. So, I take this this the this thing that this strategy that I come up with, I take the policy and I sign that and submit that to the contract, right?

To the smart contract on chain that's going to enforce the bounds and scope that I'm willing to allow my agent to execute within. And then I give the the the the rest of the strategy to the agent. It will propose high-level intents. There are a lot of systems back to the, you know, how much rule-making and protection can I have? There are a lot of systems where what it will propose is the following binary bits of an ETH transaction. Good luck figuring that out. Good luck making sure there isn't some bad idea that sneaks through or some vulnerability in a particular extra argument that was optional in some in some operation. That's not a semantic thing that a human is going to understand in terms of the policy. What you want is, yes, you can go into that

Credora vault. Yes, you can go into that gauntlet RWA vault. Yes, you can put it on base or Arbitrum or Avalanche or Solana or what have you. So, those what you want is that interaction from the agent to the smart contract to be in these straightforward to understand high-level intents. Again, it's not transfer money into account 0x426 because that's you know, easy to to be a problem. It's transfer money into my account on base that's part of my portfolio. The smart contract should know what that is. You figure it out, right?

And then the smart contract verifies that that intent matches the policy and if and only if it does, then it executes that high-level intent. Then it does the oh, transfer it over to base and put it into this particular Morpho vault, not a problem. I know how to do that. I don't need the detailed EVM transactions. What I need is the ability to the ability you know, invoke a robust transfer like CCDP and then over on base put it into the particular fund. And that's gives you the scoped authority and execution on you know, in a high-integrity fashion that we expect of applications running on chain.

Okay. But the way most systems work right now, of course, you can only do that on one chain or you have to execute off-chain which is has its own custodial challenges that haven't you haven't yet been worked through court. It has its own execution risk cuz you're running in one jurisdiction and so forth. You really want that to run on chain. This is where the orchestration approach that that the Agoric platform that we provide that that that Agoric is implemented in the Agoric L1 chain does where one contract can run an asynchronous multi-step action tracking the okay, I'm going to transfer with CCDP. That's robust. I'm going to use Axelar to send a message but not transfer the assets over to base to say do this deposit. And

so one contract gets to orchestrate all that cross-chain activity. Okay? And so that means that that mandate, that what's the scope of chains I can reach, how much how much money can move how quickly on which routes, how much am I willing to put in any particular vault or how much are we willing to put against any particular curator, those kinds of things are all implemented in one auditable place and then the secure coordination and orchestration to make that happen across all of web3 is part of the on-chain execution infrastructure.

And so that's orchestration. And what that does is it extends that enforcement of a mandate out to a trust layer that spans the entire reachable area of orchestration. And orchestration is in in many ways an important pattern. I believe how we did it is the most powerful extensible way, but there are other ways to do it. It's just a fundamental power that we need to be able to in one place say here's what I want to do and then coordinate the activity of web3 because our customers, the users out there, as we move out into the multi-router world, they don't care very much about ETH versus base versus Arbitrum. What they care about is web3 and investing in this fund or that asset.

Okay. So, as an example of this orchestration thing, we have our WiMAX product which which does this first four yield optimization. So one signature as a user's intent of here's how I want my capital allocated and it goes out and creates subsidiary accounts on all these different chains, moves the money in a reliable fashion over there so you're not sitting there with this bridge risk and you might get interrupted in the middle of this complex deployment. So one signature authorizes that, the money goes out, it gets deployed into all these all these vaults or uh DeFi services.

And so that's that's an application that's out there now, and that's the non-agentic version that gives you control with very simple interface, right? This high-level abstract rich uh uh orchestration environment, this rich cross-chain environment. Um and so uh the agentic place where this moves to is now this thing where you create your portfolio that provides that mandate. These are the funds that my thing is allowed to participate in. These are the limits on how it's allowed to trade. And then where I go is I set that mandate. So I sign the same thing. I have the same interaction with MetaMask or Rabby or or or Safe or what have you.

And submit that to the transaction. It's all non-custodial. It's all controlled by my wallet. And then what my agent does And then I delegate to my agent. I delegate to an open claw. I delegate to a Hermes. I delegate to an agent in an environment that has rich, well-curated data services inside of the YMAX ecosystem. And the agent then can immediately get to work looking at Twitter sentiment, looking at historical data, looking at current API's, looking at what have you, and propose new allocations, react to collateral starting to look a little sketchy, react to vaults being over having over demand and so they've got a liquidity crunch. All of these things to both protect your safety

and respond to issues that might compromise your your financial position and take advantage of dynamic opportunities as they appear. And so it proposes these allocations, and you know, on the side is one of the elements from our from our deployment. Um where we had to work hard to get the agent to propose one that was outside of its mandate. But it did, and it was indeed rejected on chain. Um and and so it proposes these, submits them to the smart contract on chain, that validates them against the mandate that was signed by the user, and then executes the cross-chain uh yield allocation. And one of the things that, you know, this then takes advantage of, that we discovered with the deployment of the uh the the first round of YMAX that enables you to do single signature

deployment, is that indeed cross-chain yields are better than single-chain yields, right? That on a quarter-by-quarter basis for the last many, many years, which chain has the best yields? Changes. Right? And it's not because, you know, it's sometimes it's incentive program, sometimes, oh well, you know, Polymarket was launched on Polygon, and suddenly you needed to bunch of USDC there. And Avalanche just launched an initiative the next quarter. And so suddenly you need there's a bunch of demand there. And the interest rates rise to follow the economic activity across Web3.

People want this kind of Web3 access with a level of automation that lets them not care about the details. And so that's what this achieves. Okay. And so, you know, this is currently in alpha. This is what we're working with partners. We've got some initial integrations. We've got the agent work developing. And if you are interested in participating in any of that, our booth is down over there on the left. But, let's pop up from that, cuz that's really an example of the pattern. A smart contract in one place that you submit high-level intents to, and then it validates it against the user's mandate, and executes it if it succeeds.

Right? The core of that is smart contracts. Now, let's go back, you know, the the previous thing about Cypherpunks. Well, you know, long ago when we first started doing smart contracts, what we meant by that is not simply autonomous code, but it's code that enforces the terms of a contract-like arrangement among multiple parties. Okay? And so, eBay, PayPal, Venmo, those are all smart contract businesses before blockchain. They enabled strangers to cooperate. You know, and then AMMs and perps, that's now the rich, much more dynamic, higher-speed world of smart contracts that no longer needs that trusted intermediary like like eBay in order to enable strangers to cooperate. And so that's a huge, you know,

multi-trillion-dollar solution that is working for our economy now and speeding it up and and and extending reach across the world. But there is this thing I mentioned earlier that economists are aware of that it that is exactly ironically called the principal-agent problem, which is I have an agent working for me, how do I know it's going to do the things that I want? And smart contracts now have this new relevance of enforcing that principal-agent relationship with not strangers, but rather you know, you know, you and your agents, whether they're human or computer or AI, that need to be aligned and enforcing those tools.

So, that's a you know, that's sort of a big extension of of what we're able to do. And so, with that what that's what creates the trust layer that we need to be able to run agents safely. And that's what will enable us to unleash them so that we can get to that agentic world that people are envisioning where the economy doesn't just execute at the speed of computers, but it is continually getting richer out there in a way that our agents can safely participate and safely keep let us keep up with that ever-evolving landscape.

And so, don't give your AIs keys and you'll see a lot of that when you see all of the you know, open claws and and and Hermes and so forth. Give them a mandate. Make sure that what the user wants to do is enforced in the environment so that you can safely let your agents out to play. Thank you all for your time. And uh please do check out Wymark's app at Twitter or go down to our booth over there. Um I think you'll like what the tool can do for you in terms of of uh capital deployment and yield. And um join us with the agentic building the agentic future with that.

>> [applause]