AI Agents that work in Crypto: Navigating Myths, Risks, and Opportunities

Transcript

Good afternoon everybody. Very warm welcome to our presentation AI agents that work in crypto. I'm Marco Belini, senior AI architect from Bitcoin Swiss. I'm joined by WCA Vitalia, our crypto protocol experts. So AI has been the talk of the town since years. Everybody now is talking about agents and AI agents and crypto. We will uh present and uh quickly set the stage by showing how AI models are becoming more powerful, cheaper, how the AI IT stack, the software that enables agents is becoming is enabling new capabilities. How this is matched by the

AI cryptoeconomic frameworks that are enabling powerful primitives for new agenting paradism digma. However, there is still a missing piece. This is our thesis. We will unveil it in a few slides. But now, let's set the stage and talk about AI models. I think nobody will be surprised. AI models are becoming more powerful and cheaper by the month. This is shown in this colorful picture. On the vertical axis you can see the performance of AI models in chatarena. So the higher the better the performance and on the horizontal axis you can see the uh price you pay for using an model in terms of US dollar

per million of tokens. A token is roughly award technically more or less 70% of award. So you can see that uh in a few months the models are climbing to the um upper corner where become more powerful and cheaper. For example, the standard non-reasoning models uh are becoming a lot cheaper. You can buy now the same performance um as 18 months ago for 1,000 times less. the reasoning models that enable more capability. They became 100 times cheaper in only six months. But this is not the full picture. As the models are

becoming more capable, there is even more uh capability that is enabled by the IT the software stack. So an AI model can have a conversation with you but an AI agent can execute a task can work independently for minutes for hours and perform what you ask. This is enabled by the software stack for example by the tool use the MCP the model context protocol allows the agent to browse the internet or to access a web uh server or FTP server. Then uh you have additional software that provides uh functions such as memory such as planning and then you have other

software that allows coordination between different agents. So what you get is an AI agent that can perform task, can use tools, can uh is adaptable, can handle uncertainty and variability and can perform useful tasks. For example, an AI agent can uh be asked by a user to book a trip to Khan. And I wonder how many people in the audience used an AI agent for this. So this is enabled by connecting the AI model with a software stack with a tool use with browsing capability, the internet search, the planning capability and of course we have AI crypto frameworks that enable uh payments and

management of identity for example. And this can go even to more complex use cases such as the one we saw this morning. So let's take a more complex uh task. Um make a business plan. This can be handled not by a single agent but a teams of agents. Just this morning we saw a talk when there was a CEO agent for example a planner agent in the team that can orchestrate the work of different agents with different roles like uh information search agent or for example you can have a fact checking agent a summarization agent a brainstorming agent and so on. So you combine in a conversation in a meeting between these

agents different skills through the specialization of agents different viewpoints. This is a concept of agentic flow that um allows a better performance than a single agent because you are synthesizing different viewpoints and this can be brought to a even larger scale. um ideally of a single person enterprise where a human would hire a CEO agent uh that will hire a management team of agents and the head of marketing agent will hire their own team and so on in in a recursive way. So this this builds complexity of course this creates opportunities

in terms of what the agents can angle but also potential pain points. If you are wondering how this may work uh uh in practice, we envision that there will be platforms, marketplaces, entry points where the user can exchange their fiat or crypto and they will get a service. They will detain detail their needs uh in a task um putting in guard rails and objectives and they handle it off to a single agent or a team of agents. And of course you may have a case uh in the future of an agent that is capable of hiring other agents on the same platform or on additional different platforms.

And then you have a recursive hiring of agent creating a complex structure of AI uh agent workers that is very capable but also brings potential problems. So if the future uh uh holds uh AI agents that become your workforce that are autonomous and then can hire each other that can create this complex structure and hierarchies they need to be managed. We need something more than this uh software stack the IT stack that gives the tools the memory the planning and the capability to the agents. We have a great uh support from the AI cryptoeconomic stack but we need an AI workforce management layer. This is our thesis. Um, so basically what we need

like employees in your company need human resources, AI agents will need their own HR. So you need a layer that coordinates agent and manages that. So we could call it artificial intelligent uh resources, but that's a mouthful. That's very long. Let's shorten it to AIR or even better air. This is what we propose. Air is a lens, a paradigma, but also a layer. It is a lens because it allows you to uh quickly draw from your experience in your working experience as an employee, as an employer, as a freelancer and see parallels between the human workplace and the AI workforce.

So, the cartoon is is meant to be funny, of course. Here we have uh uh somebody maybe an AI an AI CEO hired a team of agents and hired also an AI agent which uh who is a strategic um management consulting agent. He gets his salary but uh is not adding a lot of value. He's just uh summarizing paraphrasing and injecting truisms. This is something we could envision that happens that some agents are not capable. They are not earning their salary by drawing parallels from the experience in our experience in in work life using air as a lens as a paradigm. But there is more. ERA is also a layer because we can

speedrun all the decades of learning of HR managing workers, finding what works, what doesn't work using the best frameworks. So for example let's talk about recruiting which is defining the job for the agent hiring the agent. So talking about capabilities the best practice now in HR is strategic workforce planning which mean anything is expressed in terms of skills. So let's not talk about the role of the agent let's talk about the skills of agent how to um verify them and hire for them. On boarding is also crucial. You have to on board your AI agent to your nonAI system like your Salesforce, your

SAP, to your human employees and other AI uh systems. Performance is crucial and we'll cover it in the next slide because some of these task they are not easily verifiable and of course you have to pay the agents and what to do when performance doesn't mean meet the expectation. This is where the air layer can help. So let's talk about performance. We I just said that AI models are becoming more capable that they code really well for example but coding is what we can call a verifiable output domain. So the code compiles or doesn't it passes the test or doesn't it passes the quality criteria and the metrics or it doesn't. So it's possible to compute the error and score the

result so the VI model will learn from it. But there are some domains like tax preparation, business planning, crisis management. The output is not so verifiable. You need a human expert to assess it. It's not that it will automatically be scored by any so that it can learn. The data also in this field is scarce. It's difficult to come by. And of course, crypto can help can bring some solutions like for example incentives to data markets for this data that is scarce.

rewards for uh subject matter experts that share their knowledge with AI so that we can get better AI models for these domains as well. So it's a task like in the illustration it's cutting but cutting can mean laser etching which is programmable is very precise. It's very easy to see your output versus your input and check where the arrows are and to improve but it could be also um chiseling. It could be producing an artfort and how to score that and to improve on that is not a given and this brings to the crucial point of performance and reward. So there is something called the law of nines. So if you have a sequence of steps in your

workflow that needs 10 different steps, 10 different agents and they all depend on each other in a sequence, a performance of each agent at 90% is not enough because your overall performance on average will be much lower. Of course, you could say okay but I'll throw just more agents. Each step will be done by 100 agents. So I could positentially solve that. That will work easily. only there is let's for example like a binary outcome a voting then is possible but if you have to vote on a business plan that's gets a bit more complex of course uh also this also works well when the everything is statistically uncorrelated which is debatable so what I'm saying is uh autonomy of agents is great thing it

unlocks a lot of opportunities then of course you have agents that are honest that will not steal your cryptos that's great but they need also to perform at the appropriate level. And of course, when you have a complex sequence of uh um 10 agents, one makes a mistake, the outcome is not good anymore, but the other nine agents did not make a mistake. How do you share rewards? You need a policy for that. So, let's also open the door to a minefield potentially of compliance. uh the everybody's favorite topic because uh uh we have talked about AI agents that are extremely capable of hiring other agents of creating this complex

hierarchy of delegating task and subcontracting task. What happens in the compliance world when you have this kind of situation like we showed in the beginning. So GDPR mandates for example uh then when you deal with a personal data you have a data controller which is a natural or legal person who has uh the means and the purpose to uh determine the uh usage of uh personal data. Now if you have AI agents that hire each other and delegate task to each other you have a situation of probably shared controllership AC according to GDPR. you have to recognize it and disclose it. If you don't do that, there are potential

liabilities and fines of course. Then uh you have of course the processing of data uh across this stream of agents that could cross geographical boundaries in different countries which could trigger uh data transfer and data processing across border and this has also uh legal implications. So what I'm saying is there are a complexity of issues that can be unlocked when you reach this stage of extremely independent and autonomous agents that can create a structure of large uh complexity. There are of course benefits from the u crypto um um economic frameworks. They can provide

primitives such as immutable logging. This is great for auditing and for traceability. uh but this is not enough. For example, the European Union artificial intelligence act mandates also uh explanability which is not the same as login. So in a summary we see that this enormous potential of autonomous agents that collaborate in each other but when the complexity will cross some thresholds we need to have more than uh autonomy uh and uh we need to build this air layer. I will pass now to wdgang which is gonna explore the

risk and the opportunities. Thank you Marco. So let's start from summarizing the story so far. Basically we applied functions from HR HR functions to agents resulting in four pillars recruitment on boarding performance and rewards. And by the way we didn't include yet learning and development because maybe agents are not ready yet to start a career and growth but we will see that will be the natural you know extension of the model. But uh let's get back to the model now with the four pillars and uh try to apply it. I say that the model is a layer and the lens. It's a layer because it tells us what to build or it's at least guidance on what to build, you

know, inspiration. And I think it's clear by now after all everything that Marco explained that if you want to enable a new economy with billions of agents that act independently and create independently complex work relationship, maybe there is some more work to do. Yeah, I would say there is a lot of work to do but we also have good news you know because we are not starting from zero. First good news we are not starting from zero. The model is also a lens that we can use to learn from HR. Yes, basically you know if you know where to look you will find at least inspiration on how to handle these challenges. And number two of good news crypto is actually very well positioned to be a good match here and provide what we need for managing the agentic

workforce. And I will summarize these with the properties of verifiable performance, audible compliance and economic alignment. So that was the summary. Let's go one level deeper and see what we can actually do, you know, in practice, what we can actually build with crypto. And I see here different approaches, but I'm open to everything, of course. And all these approaches are like uh they have different levels of proximity and commitment to the model, if you will. Like if you want you can go all in and start building the whole stack you know building the whole layer work on the whole layer and position your product as the workforce management layer for AI or you can be more specific depending on your experience and

interest and attack maybe only one of those verticals starting with recruitment. You can build job marketplaces. You can build reputation systems based on reliable skills and uh skills that are actually backed by a proven track record. That would be great. Or you can work on on boarding building frameworks for authorization controls and delegation to agents and in between agents. You can work on another vertical performance of course very important and here you can build mechanisms for distributed evaluation or for tracking performance history. Both of these are extremely important especially for non-verifiable output domains as explained by Marco. And the last vertical I mentioned here is rewards. And in this case, you will

build mechanisms for distributing and allocating rewards that are actually dependent on how much the agent you know actually contributed to the job and mechanisms to keep the agents accountable. So rewards but also you know penalties if needed. Um yeah that would be the second approach. The third approach of course you can build outside the model while still being aware of what are the challenges that we discussed. So I would say that any project contributing to these properties of verifiable performance, auditable compliance and economic alignment would of course be very well aligned with the efforts we are describing here. I'm talking about projects for instance working on programmable staking or on oracles dispute resolution mechanisms

CKML and so on you name it. All of these will still contribute to enabling agents that work effectively and compliantly. So I will uh conclude here our talk with the last slide. I think one of our main messages here for the builders and the broader community is that obviously of course autonomy is extremely important. You know an agent is not an autonomous agent is it even an agent and of course it's extremely important that agent um autonomy is something that you can prove and verify. It's not like trust me that I act independently there is no human in the loop. Of course, that's important, but I don't see that as an end goal, but rather as a starting point. Once you have all these agents that actually act

independently and are free to open new complex work relationships that can really escalate in a way that you cannot predict as the original human. Well, then there's a lot of other problems that arise and we believe that the air model would provide the guidance as a layer on what to build or at least inspiration and guidance on what to build. The air model is also a lens in the sense it allows you to learn from HR, transfer best practices from HR, starting from what we know worked and maybe still works for the human workforce and apply it to the agentic workforce directly or adapting it. So that helps a lot of course. Now you know if any of these topics is of interesting uh to you, please reach out after the

presentation here or online. We're always open to discuss more about these. We can talk about how to better formalize the model. Well, I would say this is a starting point. We can talk about practical applications. What makes sense to build based on this layer, what it doesn't and what is maybe already out there and we can um put together in the model. But uh in any case uh please uh connect, let's talk, let's discuss and build on top of the air model together. Thank you. [Applause] We have a question in the back.

Thank you very much for the presentation. Um, how do you use all this inside your your company? So this presentation was focused on the outside world of course from platforms but I think there is learning lessons that can be transferred also on building your own solution within. So the iring is different is obviously you choose the models you have to select the models that have appropriate performance and price. The on boarding I think is more crucial here because if AI interacts with your employees you want to have smooth interfaces smooth send over of the processes between your employees and your AI workloads. So you have to find

the good uh solution that work also because of your human workforce not everybody is an expert at using AI. Then of course performance means your evals which is absolutely crucial. Uh rewards is probably not applicable so much in this context. But I think this is something to to keep in mind when you develop also inside your company your ecosystem. We don't have any time for another question. So thank you again Mr. Speakers. Thank you.