Panel: AI Agents in Crypto: Lessons from Phase 1 | ETHSofia 2025

[music] So um let's start with uh a little introduction. Uh we are moving faster than ever. Um and uh AI agents are all over the place. Uh automation has never been easier amongst the space. uh companies are laying off staff to experiment with agents.

Uh we are seeing other uh agents trying to fulfill the need for people who understand uh the topic and are capable of implementing it within the within the space. We are seeing uh people quitting jobs to become solopreneurs building their own uh AI agent uh empire let's say. So I am honored to sit amongst such experts uh and uh talk a little on the topic. Um so let's uh begin with our first question and uh it's to uh Daniela from uh your vantage point.

Hello. uh from your vantage point uh in banking and AI uh can you tell us uh how do you view uh AI agents and how do they relate to LLM and uh of course how is banking currently uh benefiting agents

yes of course right so agents are not something new uh they if you look at a very famous book which was issued 1995 uh there is already a definition of agents but they were not popular in the AI world and especially in the business but the rise of LLM make them possible and also made them needed. So with LLMs we get a reactive behavior. So we send a request we get a response but then what we need in financial institution and in any other enterprise or even smaller businesses is a more proactive behavior and this is what the agents are giving us. They are giving us a way to um make them autonomously uh working thinking using our already established infrastructure and services and even store information and have kind of um memory like humans do. So in financial institutions and any other business of course they will be the future for sure because they will provide us the needed um gap for additional automation of um manual processes and also they will help us in coding and all other stuff that is happening in banks and even in software companies.

Great. Thank you very much. Uh so uh Titos from uh your point of view from uh your perspective working with customers uh in AI how much awareness is uh there and uh how much demand is for AI agents at the moment?

So I think everyone building in uh any kind of technology space right now wants to implement some form of AI in their business. I think there's a lot of conversation happening around the fully autonomous 100% automation. I think that's generally a wrong initial approach. Um people who will be the most successful in our space will be the ones who will learn how to augment themselves and how to augment their teams with AI and not go fully uh autonomous immediately. We saw some of those initial lessons in uh in crypto agents uh version one where we had those like Xbots and then all these like autonomous trading tools.

they work if you have some expertise behind uh but then more often than not they their errors tend to compound if they're left unchecked. So I think we need to educate both ourselves and like the people we work with in the space to understand like what kind of guards we need to put in place, what kind of controls and the way I tend to think about AI agents is like it's more an extension of my own capabilities and my team's capabilities or the client's team's capabilities as almost like an Iron Man suit as opposed to a fully autonomous like R2-D2 or C3P or robot if you want to kind of have the movie analogy.

Uh great, thank you. And uh Evo, would you say it would you say that u the hype that has been around uh for a while related to uh crypto agents is uh something that can uh finally uh be justified uh and how important are agents uh in this regards?

So yeah um that's a fantastic question. M so uh first of all you have to keep in mind that AI plus blockchain has been something sort of promoted and hyped since uh 2017 but mostly it has always been vap vaporware and it has never really made sense because AI and blockchain are fundamentally like almost opposites right because AI can never run on blockchain and this was never the right mental model to begin with so blockchain has really limit limited computational capacity and AI is all about having as much computational capacity as possible. So they it sounds like they don't make sense together. But in reality, what ended up happening is that AI agents are finally the thing that makes AI plus blockchain make sense. Why?

Exactly. Because they're opposites. Because blockchain basically can put in the guard rails to because are very um nondeterministic. They're probabilistic, so they're kind of random systems. They're unpredictable and they also hallucinate a lot.

So if you want to put them to use for financial applications, which also uh sounds like it makes a lot of sense because they can make autonomous decisions. But if you want to make them uh make financial decisions, you need to put in guard rails, which only blockchain can really do because you can write a smart contract which puts in some set of rules on what the AI agent can and cannot do. And then you can feel safe about giving an LOM control over to some funds or some important operation because as as I said Loom hallucinate they're not perfect. They're uh probabilistic. Um so you sometimes cannot predict their output.

Sometimes they would just uh give you like um incorrect stuff. So putting those guardrails on chain finally ma finally makes AI plus blockchain make sense. So it's sort of like an yin and yang thing where uh one is the security and the other one is the creativity. So one is like deeply grounded which is the smart contract the blockchain and the other one is creative or like has some interesting outputs. So yeah I think the hype finally after all these years makes sense.

So we should be happy that we are in this space then.

Absolutely. Uh okay. Uh Titos uh let's expand on this topic. Uh from your experience, what needs to happen for agents to go uh beyond this early adoption phase? Uh is it about lowering complexity?

Is it about better tooling? Is it about uh building trust or is it something else?

I think it's all of the above. And I think I would start with uh better tooling because that then gives you the leverage to implement more use cases, experiment with more scenarios, more workflows, more toolkits and then that builds that trust that that you need to essentially scale and give it more autonomy. I think we're also getting to a point where guardrails become interesting. Um, as Evo said, uh, you can independently verify those if they're coded inside the smart contract and there's proof on chain in metadata, for example, that the execution actually happened. Um, and, uh, now with, uh, payment infrastructure as well being set up with X42 and then with AP2, uh, two new protocols, well, relatively new protocols, but in this space, if anything that's less than a year old is considered new.

um with those I think there's there's more ways to um create autonomy within a verifiable space because what you want to do is you want to check whether an agent actually made a payment before releasing certain like data or uh reward or allowing to enter a certain space or like clearing a trade for example but you you can do that now autonomously if you look on Jane uh for a transaction specifically tied to that action right with the explore 2 protocol uh and I think That's that's where we're building that muscle to understand how can we do first 10 of these, first thousand of these, first 100 thousand of these and then millions of transactions and then that essentially adds a layer of uh Aentai to right now one of the biggest use cases for crypto which is DeFi.

Yes. Uh Vivo, do you want to add something to the topic? uh yes I can basically make a slight transition uh with and taking the tooling as a keyword that you that kit mentioned and make it a bit broader it's uh infrastructure and architecture so for now we had this microservices monoliths and and eventbased systems and many more but we didn't have identic infrastructures and in general uh our infrastructures are not ready for that in order to adopt uh those systems of course we can implement very easily even with a no code to an agent but how that would work in in production we don't know because even the big players like Google and Asia even they are not prepared with the right tooling for agents in terms of security in terms of um tool registries agent registries agent communication especially data protection and so on so this is um something that is about to come. It's something that even the big players and banks and all other institution are on their way but we are still not there. Um I don't have much to add but it's worth noting here that we need to maybe we needed an introduction on what agents are in the first place and um basically u I would personally define agents as like a system where anom interacts with the outside world uh and has some sort of a set of rules or a framework where it can loop.

This is the most important thing about an agent. It can take stuff from the environment, analyze stuff from the environment, then affect the environment and then continue operating on this loop of learning what how it affected the environment in some way. So

reasoning.

Yeah, this is well you could argue that it's not reasoning. It's just like interaction like your senses like your eyes, your ears etc. But what I want to remind everyone for the context of this conversation is that CHP itself is an agent. It's no longer just an aom like the very early versions I think they were almost a full no there were still kind of agents because when you type something in charge um it actually you don't see the entire response part of the response is its memory and then the next time you type something into chargept it sends the whole memory of the entire conversation that you have and the things that it was supposed to remember so it's already a loop and now in these days it also searches Google searches the internet etc. So, it's literally an agent by this point.

And a true agent, like even more agentic than that, would be something that interacts with the blockchain or something that interacts with the world in general and it can create an account for you or whatever. But my point here is is that okay, we're not that far ahead in terms of infrastructure, but a lot of

radio also. Yes,

not really. But at the same time, there's a lot of stuff already built and there's a lot you can already get quite far ahead with the existing stuff we have. The biggest challenge is actually that you are absolutely right. We are much ahead in terms of development of agents but not the infrastructure and there is this gap between many teams are implementing agentic system. Many companies even small companies startups are implementing identic systems but the whole infrastructure is not matured yet and that's the biggest challenge and something that we also discussed um maybe you will touch it but I'll just um make a point on it is this knowledge gap of teams where everyone thinks he understands AI and agents and everything but people are still not uh that far in terms of training and knowledge and experience.

else.

Yeah, 100% agree. And I I wanted to go back to like the infrastructure uh gap. I think it's very tied to the knowledge gap and I think there's a lot of infrastructure for developers out there, right? For people who understand the technical underlyings of how you need to execute a autonomous system be it like service inside of your organization there's well-known like AM RBAC for that. Um you can apply a lot of that to Agentic AI.

I mean in a sense it's still a program. It still can function as a service account. But the gap between like a developer deploying an agent inside an organization and a non-developer and people who are not as technical deploying uh agents organizations is closing. And that's actually why we launched Nexus when we did the yesterday is to give more people power with uh less technical capabilities to be able to have that access inside the organization because again coming back to the Iron Man suit analogy that I did earlier if you are able to embody the agent with the access that you have inside an organization like be it you're a sales uh representative be you're a marketing executive you're an operations person you have access to certain data and certain accounts And as long as you operate within that guardrail and the agent operates within that guardrail, then you can do um a lot more work faster with the added intelligence capacity that the LM brings in because uh and coming back to Evo's uh conversation around Chad GPT and be how it was initially a closed LM, it started becoming an agent when it got access to the web and it started getting access to tooling. But now I think the question is like how do we get it outside of the um IT circle of fear and more into hey now we're creating a distrusted execution environment where we can operate with our own data like we do day-to-day in our in our work environment because as as employees we have access to data to do our work right and then the same way now we bring this LM be it like an in-house one or if you're using a third party one u you can uh add that layer of intelligence to your uh your day-to-day Uh great.

So let's let's uh park our discussion for a second. I want to ask uh our audience how many of you guys have actually built an AI agent yourself? Raise your hand. Okay, we have like at least 10 people. Great.

Uh great.

Maybe we can also ask how many of you would like to do agents? Yes,

not that many.

Um, so we uh already identified one gap uh which is the the knowledge gap considering that agents hallucinate governance is tricky. Uh security is mandatory. Is there um something else which we can define as uh as a hard truth that we learned from phase one? something more to add to that?

I think demos are fun and production is hard. I think that's a super hardcore fact that everybody who's doing AI learns. And then the other one I think in my in in my experience is you need initially to be the human in the loop when it comes to deploying your first few agents until you learn how to operate with it and what it can or cannot do. Then the other one I would say if you're trying to automate something you don't understand very well then you're going to fail.

Okay. Uh yeah

yeah I I think that Tito already mentioned it in his previous comment like we identified the knowledge gap but he mentioned something much more important also that uh challenge is to enable people to use that. I mean the enablers are pretty much the people that need to make it easy and straightforward in order to spread it and make it useful and that's why many companies right now are failing with doing that because there there is a need of key people that uh are able to enable other people to do that job. And another challenge of course is security that we need to um deal with because we're talking about financial institutions here. So security auditing um banking regulations are key requirements and they need to somehow fit in that um picture. So, um I'm not sure if those are hard truths, but I would like to give some advice to everyone who aspires to build an a an AI agent.

And it's kind of a hard truth, but it's don't get fixated on doing everything through an OM. Like sometimes an old-fashioned if else is best. And when talking about agentic systems, they're um there this composition of multiple components. So in an agentic system sometimes you can have multiple Oms interacting with each other. Uh sometimes you can have an LOM and some check some verification a smart contract whatever but my point is don't get fixated on running every single thing through the LOM and some things are really um a simple traditional code can do them better like for example an LM cannot beat a chess algorithm at chess.

This is something that everyone needs to like keep in mind all the time. Uh so sometimes an if else or just an algorithm is better and make sure to compose the system in a way that takes the best advantage of every specific thing for the specific goal and don't overuse the LOM. I I think evil statements is really valid in terms of don't use for everything LLM and it's absolutely valid also for a agent and uh sometimes it's it's enough to use an LLM for something small instead of building a whole agent. Uh a simple example would be we all know there are already existing business processes which are automated. So simple steps might be just candidates for an LLM.

later on it could also uh be transformed into an agent but starting small is much safer than just going to this uh complex agentic systems.

I I have a bit of a contrarian approach on that end. Sometimes I found that it's better to use the agentic approach and the LLM to find the patterns in the thing that you're trying to automate and then from those let's say you find that 30% 40% 90% of that can be turned into something of a mechanical u approach. Uh and the best example I can give for that is data scraping. Initially you don't know what you're scraping. You need to find structure in that website for example.

you strip the HTML, you find the objects in there and then at some point you realize there's enough repeatability because the the website won't change as much uh so that you can build a mechanical extractor and that's incredibly more efficient both from a time and from a cost perspective because let's not forget AI agents are very expensive at scale. uh those old tokens, all the compute, all the memory that you need to use is not trivial. Whereas a mechanical approach usually means that your cost will be 10, 100 or thousand times uh less. But to get there, I tend to uh like appreciate the help of the LLM to analyze that data and get there faster, right? Find the patterns, automate, and then use the agent on like new fuzzy things that I want to discover.

Okay. I have a a question regarding that. Uh do you think that this will make us too dependent on AI going forward? Uh I'm I'm talking about my perspective from my own experience. Uh the more I use uh AI to advise me what to do, the more dependent I become.

uh the less I think myself and uh then there is a moment that uh happens and I'm okay I need to start thinking myself and making decisions myself and not rely on LLM to to to uh to make the decision for me or uh do you have a filter like you use the AI uh to suggest some actions and then you filter them out consider what's the best approach.

I mean I I read about this I think just yesterday. There was an example of if you outsource your thinking to AI completely then you miss uh the ability to understand your process and you discount your own expertise in that field. And I think if you want to get great results from an agentic system, you need to do it in an area where you have some expertise because you won't be able to tell the good answers from the bad answers if you don't really have expertise in the field. And like with programming, you if you're not a experienced developer in a specific uh environment, you don't really know what's a good code, what's bad code, what's the security risk, what is a good component, a bad component, an efficient use of use effect in JavaScript or an inefficient one. uh it why is my code crashing if I have more than 10 users on my application at the same time right so those are I think the the things I think about when uh becoming dependent on LMS or not I think we already are to an extent in the same way we're dependent on the internet or on our phones or on our like day-to-day use of electricity so in a way it's like adds to the stack of things that we have to improve our lives be it work or home

great thank you for that answer uh Evo uh so looking

can I interrupt I have a hot take about this.

Yes.

So, um the thing is yes, AI makes people lazy. All right. And some people tend to stop thinking, but uh is it really that much different than Google before? And is it really that much different than like looking up Kora or Reddit or whatever like user generated answer is to something. Um I think it's not that different.

Yeah, it makes people a little bit lazier, but in reality, it's just an extension to what we had before. And especially with coding you can see this because even before code we had copilot and before that we had autocomplete and some autocompletes were pretty smart. So like coding these days is just a little bit more of that. You still need to think you still need to apply like human engineering and like this mentality of doing things the correct way. It's just a little bit like more automated.

Very interesting take. Yeah. Thank you. Uh then let's proceed to that question. uh looking at uh what we've learned so far, what do you think is uh the glue that will carry us from phase one to phase two and uh what bridges need to be built for agents to truly deliver?

Fantastic question and um I'm going to introduce a shameless plug a little bit later, but um let's start with this. What do we have right now when it comes to AI agents and web three? What do we have in terms of this intersection? I would argue that we don't really have anything. So almost all the systems that exist are really rudimentary and they're just a wrapper around JBT often or a wrapper around another and they're all always um prompt in in prompt induced systems uh where the user has to chat with the M in order to get some results.

I think the next step uh would be promptless agents that work in favor of the user. So this is something that we are building at uh addex and we're building an AI agent framework for basically for for your wallet and this is something that we are going to integrate with umpire. So the vision there is that um the user would simply enable this agent once and this agent would work in the background to claim airdrops uh to like invest on your behalf safely uh to like optimize your portfolio like if you have stable coins lying around it will put them in a or wherever the highest yield generation is. But I think this is the big step from stage one to stage two. Getting not exactly rid of prompts, but building systems that don't actually require prompts to work because let's face it, prompts are kind of we're already bored of them.

Like we all know charge. We all know it's useful, but we just use it as another Google or sometimes we use it for deeper research, but it's not it's not what we think as an agent. I said it's an agent. Technically, it's an agent, but it's not what it's not the future. It's not like something that works in the background on your behalf and actually delivers value to you without you even knowing.

Basically, this is exactly what agents by definition are. They are perceiving their environment. They are not waiting for someone to prompt them.

Exactly. Exactly. Even though Chip is an agent because it still perceives the environment, but it does this after you prompt it. So, it's like a mini agent and a shortlived agent. a real agent would operate all the time autonomously on your behalf.

I I think I have again a contrarian approach to that. Um I think prompting should be boring in the same way investing should be boring for most people because it produces a predictable result. And I think we need to have a lot of prompting that is super boring but does the job. And I think we need to have also these autonomous agents that go at the cutting edge and operate autonomously and bring us value because ultimately also if you think about prompt if you give a a prompt that needs a longer running cycle that's almost like you're spinning up a temporary agent and I think we need to get used to working with these temporary agents more and more and giving it more and more trust to get to that point where you have the well there's enough habit or history learned that now the agentic system knows me or knows my my goals well enough or has borrowed from my goals enough so that the result is compelling. Right?

For example, Chad GPT, it has history from uh 2012, 2012, 2022. For me, it hasn't been that long. Um, so I've been using it for many years. And right now, when I ask it like, hey, uh, help me write this article with these topics in my style. That's all I need to say because it has enough memory of me and my goals to understand like how long does it need to spend thinking on that topic gather information and what kind of information I want how do I I want it to site for me in the same way it can do for for other tasks like gather data perceive the environment in some way gather uh information to get me for example in an autonomous vehicle uh use case from point A to point B without crashing or stalling or getting blocked.

I think you you both guys can meet in the middle because there are also systems which are with prompts but they are optimized prompts like prompt engineering, finetuning and so on. So they're very much systems like uh right now they have a predefined set of prompts which are used behind the scenes and simulate kind of autonomous or fully autonomous behavior but still based on prompts. So that's a step between you and uh between you in terms of um identic systems.

By the way, um don't get me wrong, I'm not against prompts. What I'm saying is that the question was how do we get from stage one to stage two and prompts work and they will continue to work. What I'm saying is that stage two maybe involves less prompting and more automation. Yeah. And maybe more more behind the scenes management of that uh prompt.

I agree. And then we have MCP the model context protocol that's helping like standardize how we bring data into context in the same way like prompts can be standardized with system prompts with developer prompts behind the scenes. So don't you the user doesn't have to spend I know hours building the perfect prompt but it expresses an intention and then gets the result.

Yeah. Which is by the way super aentic but shortlived.

Yeah. Yeah.

Nice. Uh okay. So uh we are running a bit out of time. Uh let's go to our last question and then we'll have few minutes for questions from the audience. If we're sitting here again in 10 years, it's not that long.

What headline about AI agents do you hope we'll be discussing? Or put differently, are agents going to change the world in a similar way to how the internet did?

I think it already started. um in 10 years I think we're going to have companies 10 times or 100 times more companies with the same amount of people doing work at a scale we can't even imagine from a data perspective u because we now have access to almost like perfect in perfect perfect inference perfect intelligence combined with total data we can do um well we can solve much more problems at scale I think that's a super high level definition and I think at we'll have an AI agent panel at that point with us on top of us.

Uh you know I'm in that IT business for more than 20 years and during the last maybe five years I'm constantly asking myself the question how is that possible that nothing changed over the last 20 years and suddenly it happens. So now things are changing and the only thing I can tell definitely yes and I would allow myself to state Huang's the CEO of Nvidia statement on his presentation. Yes, LLMs and agents are the next industrial revolution. Definitely. Yes.

I'm not sure we're going to be speaking about AI agents in 10 years. Let let me explain. So once a technology becomes successful, it becomes so embedded in our toolbox that we forget it exists like are we talking about React frameworks? We're not. Um obviously at a React conference they would be but we are at a web three conference and we are talking about AI agents and everyone is talking about AI agents at any conference.

Um so I think that once a technology proves itself enough it starts getting taken for granted and it starts getting it's something that exists uh people work on it people develop it but it's no longer like this magic so uh I think in 10 years AI agents are going to be there and we are no longer going to be discussing them as agentic systems because in reality if you drop the if you forget about a gentic systems already exist exist like Google has a scraper uh it's even machine learning learning uh powered and it's always existed like for the last 20 years Amazon also. So all of those big companies are really complicated automated systems and no one is talking about the machine learning behind them. Um people were when Google was hot uh when it was emerging but they're no longer discussing this. So I think it's kind of as much it is a revolution it's also that much of an evolution.

Nobody mentioned AGI. Uh why uh if if I have to say like a title for for a panel in 10 years, it would be something like uh AGI is here. What's next? Uh what do you think? Uh uh will will we have AGI by 10 years?

If you've ever tried to integrate multiple data sources in a single context for LLMs, you know how far AGI is from us. Like as long as we still have to handle O tokens, refreshes and then connect data systems and match do uh matching tables between data structures or at least put an LLM to do that. We still have a long way to do before any AGI comes to market. I I I'm really afraid it's not directly related to the topic, but what I'm afraid that right now we are discussing how to make agents smarter and I really hope that in 10 years we will not be discussing how to make people smarter because of AI agents. So we need to be careful.

No comment on AGI. Maybe it already exists but okay. Anyway, um I really agree with Titus because if you've ever really tried to build something with an M, you will hit the limits really really quickly. Like if you try to do something more complicated than something that someone has already done. If you try to innovate, you will see very quickly how dumb this system actually is and you will realize like how difficult it is and why it's so valuable and why uh billion dollar companies get created in that way to actually make an app anom appear smart.

So you will realize that charge is a marvel of engineering not only the AOM behind it but also the system that it is and you will realize um how difficult it is to make this system appear as smart as it it appears in the big products like JPT. Um so I I don't want to say whether we're far from AGI or not because the current system is really dumb in comparison to like full human intelligence. Uh but in many ways like the smok and mirrors illusion is working and uh in reality when you think about the brain the brain is actually like arguably a deterministic system but it's a physical system and biological system that should be replicatable in theory. So why not?

Okay. Uh thank you for the insightful discussion guys. Uh glad to be among

thank you for the moderation.

Thank you. Heat. [applause] [music] Heat. [music]

[music]