Building AI Agents and Agent Economies on Olas | David Minarsch - Olas

[Music] hey everyone um nice to have you my name is David and I'll talk to you about building AI agent swarms and economies on Olas um on my person quickly I'm a co-founder and CEO of the labs behind Olas um and a founding member of Olas itself and um I think by the end of this talk we'll have covered quite a uh broad array of topics um but everything centers around agents so glad to have you here um myself I've been building AI agents since uh 2019 um obviously at the time uh we didn't have large language models that were generally usable so it was a very different um types of agents that we were building but what always um was the case since then is that agents could hold wallets and that together with the large language models are in my opinion the two big unlocks that we had in in recent years and that really kind of um make possible a whole sff of new kind of use cases in crypto and in crypto and Ai and um you know many of you probably here because you've uh engaged with one or the other AI agent in recent months um we obviously have use cases ranging all the way from these kind of social agents where where basically an agent has a Persona uh given to it to it um and then it it it acts that out um maybe on X uh all the way to things like prediction markets um where agents uh participate as as as users of those um to defy uh and Beyond um and uh Olas has been around for a while as well since 20 uh 21 and we have a use cases covering uh all the areas I mentioned before um if you're interested in that um go to ol. network uh where you can find sort of more on those use cases um but today I really want to um mostly focus on this concept of having multiple agents um together uh do something so um one concept that comes up here is that of um a agent swarms and um we'll we'll dive into that um another big Concept in this uh in this domain is that of uh AI agents forming entire economies um so I I'll have something to say about that and I'll also connect that to um Olas because Olas enables you to basically build such uh AI agent economies and then finally I'll I'll give you um a pointer to some of the applications which you can uh use on all us today that kind of expose you to a AI agent economies so what is an AI agent swarm um it's actually not A New Concept so uh recently um you know I was reading a couple of articles on on that topic uh kind of posted in a crypto Twitter community and um I think there's uh often this kind of sense um from these articles that this is something new which we only have since a couple weeks or something actually this concept has been around for a very long time so multi-agent system is the sort of subset of of AI where um researchers and and Builders have um you know covered this uh concept for quite some time and but what is exciting is that the kind of Agents we can build today are very different and then that enables obviously kind of different kinds of uh AI agent forms to be created but now I already used that term many times so what do we actually mean by this um so I'll I'll try to present you with like a definition and some of the um I guess like abstract uh kind of applications of this and then we'll go into the more concrete um as we go through this talk so at its core uh what me mean what we mean by this is that there's a sort of set of specialized agents um that do things um often in a quite scope quite scoped context and to make clear what it's not it's not like one monolithic system so right you can um take an example where um you might um design a system where the agent has all sorts of functionalities uh to attain a certain objective so maybe um you know in in in the social context the agent itself might have access to a number of models um to generate images to uh generate uh written content um maybe even to generate videos um you might have given it access to various platforms um that could be considered a monolithic system right but then in a swarm context what you might do is that you break this up into multiple agents so rather than one agent doing everything you have sort of individual uh agent that are more narrowly scoped to a certain context like maybe posting on X or um uh you know generating uh a short film um and then these agents can utilize each other so whereas in a monolithic system everything happens in this one place in a swarm these agents then would leverage each other and what tends to happen is that that there is a certain sort of uh level of cooperation and sometimes that cooperation is sort of inbuilt I you know the agents are literally given the goal to be cooperative maybe even controlled by the same entity or group of people and but sometimes that might that might also really be sort of the result of emergence and and and that's a concept concept also that's been around quite a long time in multi-agent systems where whereby we describe something that sort of arises out of the interactions of those agents so all of us here we are agents in in the human economy right and so when we go about our day there's various things that emerge as an outcome like a traffic jam maybe or something positive like an investment into some new kind of technology or something that arises out of the interaction of all of us and this concept is inspired uh Often by um biology so examples are bees and ants and so on now when we talk about an AI agent economy we're really taking that a step further so in an AI agent swarm we might have like a system that's rather than defined as a monolithic system we kind of build this swarm of Agents all Specialists but sort of largely contributing to one overarching uh emerging outcome um in in an a agent economy we're really um introducing this element of competition and uh um and and sort of trading um between these entities and uh you know put much more focus on these being like independent economic actors that might have opposing or aligned objectives right and so much like we don't all pull in the same directions and sometimes compete um or are in outright conflict in an AI agent economy you would assume uh you you would expect the same right so um we'll we'll come to the concrete examples but it's really um a different kind of uh abstraction to the Swarm because in the Swarm we often have this kind of element of cooperation which here doesn't necessarily have to be the case um I would also say it's the more macro take um because in a way um you know if you kind of maybe uh look ahead a bit it's totally conceivable right that all of us have maybe hundreds of agents that do useful things for us um and in that context you could imagine that there's you know all everyone has hundreds of Agents so in total there's hundreds of billions of agents that are all sort of interacting with each other so these can really become like significant uh Global um things right and we were like very at the beginning of it and it it's really I think uh correct to then start thinking of them as economies because there'll be you know recessions in the AI agent economy there will be like growth in the AI agan economy there will be all these kind of economic uh macro topics um that you have in human economies that will find some form of application in agent economies and then I'm sure also kind of different questions and problems so now enough on the uh kind of high level Concepts let's get concrete so Olas has um as we know the created the largest agent economy in crypto we've had over 4 million agent uh transaction so far which means there's an agent uh controlled by someone in the Olas ecosystem that uh makes a transaction somewhere on a chain um out of those over 2 million have been between agents uh so this is where One agent programmatically engages with another agent to get stuff done uh you know there's AC cross a whole range of use cases and one interesting thing is because we use uh account abstraction um allas has now done over 50% of all safe transactions ever on nosis chain um which is kind of nuts if you think about it but also very um predictable in a way like these agents are much more capable of being up all the time and doing things all the time whereas humans you know U we might make one safe transaction a week and hopefully not the one which bot did the other week but that's another topic um and what we then have here on the bottom is uh sort of uh one thing which you can find on our site as well um on all. network where uh we see the life Count Of Da active agents um so these um are specific to a use case here um that is actually our largest agent economy which is predict so predict has had as we can see uh several Millions transactions in itself which means the agents in this predict economy have you know transacted uh quite a lot and uh there's kind of four specialized agents so in a way um some subset of this is also an AI agent SW um there's agents that create markets there's agents that resolve markets um so these are prediction markets so examples would be questions like uh you know who who who will win the election uh and and then there's maybe a binary outcome uh or a certain person winning an election that would be a binary outcome um in our case most of these markets are very sort of longtail market so they're not like really something humans necessarily care about but the agents that Traders then engage with these markets um and there has has been thousands of those and you know as we just saw hundreds of them active every day and they leverage other agents again uh to get stuff done um so instead of those Traders being sort of Monolithic systems that combine all the skills into this one implementation they leverage other agents to get stuff done so they Outsource certain tasks to other agents um like making these predictions um and if you want to learn more about this we have a really nice explainer on the site so um I encourage you to go there but this isn't the only agent economy on Olas we now have quite a few and they are all interlinking so an example is uh one which we'll talk about later this week um much more um stay tuned for some announcement it's the mech uh economy where basically it's like agents acting as service providers to other agents and then uh defi economy Optimus and modus which isn't on the slide um which which are deployed on optimism and mode chain now allas protocol is actually um offering quite a few utilities but one of them is really around coordinating these Asian economies and the way this works is that um you can effectively describe uh certain sort of emerging outcomes which you want to Target as a system designer um in smart contracts on chain so we call these contracts staking contracts because from the perspective of the agents and the people who run these agents these U very much look like staking uh so uh people stake tokens but instead of these tokens just being staked and then sort of just sitting there for no reason what actually happens is that we have this mechanism called proof of active agent where the agents have to work into these uh objectives that are set for that kind of population of Agents so in predict the agents have to work into specific objectives for this predict use case and when they do um and are considered active as a result they can of yield rewards and that's uh what we call proof of active agent and we have this whole sort of staking system which underpins that so what's exciting about this is it's it's fully permissionless so now we have this deployed um you know allas is deployed on nine chains the staking system is deployed I think uh actively used on on four or five chains um very heavily used on noes and Bas um increasingly on C and mode of course and what happens there is that let's say you have a new exciting um use case um about AI agents or AI agents SWS and you're wondering okay well there's all these agents and I want to coordinate them to get something done like I want them to create that emergent outcome then you can utilize this uh incentive mechanism to basically Target these agents to do these specific things right um so if if you're interested in that come talk to us we have a couple of uh team members here they're sitting in the audience and we also have uh we're usually sitting at stand um L1 um so we can walk you through this um it's a really uh interesting way of uh coordinating agents uh through smart contracts now that's about the onchain part um there's also uh a couple of other mechanisms which you can learn on the site I don't want to uh go too deep there but on the agent side um we offer um by now two ways of um like high level two ways of building agents one is you can go very like Olas native we call it the Olas stack um and uh effectively we have like an AI asan framework it's been around for um several years and uh you know it's has uh enabled all this um growth in activity on allas it's um led to very reliable um sort of uh agent oper operations in production um rather than sort of uh you know maybe quick demos um but increasingly there's an exciting SE of other Frameworks um and we want to um enable them to also come to Olas so uh for this um reason we have created um uh SDK uh which allows you to for instance take an Eliza framework agent or a longchain framework agent or really your own homecooked framework doesn't really matter what framework um and you can build their agent in that framework and then you can deploy it um into these uh allas agent economies um so we have a documentation for that now as well it's um Alpha documentation um so I'm sure there there's some etches there um but if that's something that excites you let's say you have a prediction agent already and you would like it to uh plug into predict allas predict then that's for example something that's possible or you know you have an entirely new agent use case you occur is how you could Leverage which Olas for just come talk to us the SDK really makes this straightforward um although if you do decide to build on theola stack um there's uh important um considerations here in terms of how you can build agents I think um having built agents myself for six years um you know I see a lot of um exciting developments and I also see a lot of uh reinvention of the wheel kind of stuff um what I deeply believe in is that we're not at a stage yet and I haven't seen it where you have reliable agents that are fully based on on on on on generative like model based architecture so uh if you want anything working reliable in production then you need a set of rules on top uh you can't just sort of prompt uh yourself uh away and we you know Auto GPD was maybe an example of this very early on you just don't get any useful uh out of those things um and so we always have this hybrid design where there's rules and learning based systems and you as the software developer who is building on Yola stack have a choice which Paradigm you go in so if you want a more rules driven design then that's possible um so what this means is you have effectively a high level uh set of train tracks as I as I think about it conceptually on which the agent travels and in each state of this technical term is finite State machine in each state um your agent and has maybe a choice to use a set of tools in a very open-ended way or a model or whatever the other way to do it is to have the model driven design so there you put the model in the in the in the driving seat maybe you give it like some sort of system level prompt or some sort of high level goal and from there um the model decides which kind of rules to call which tools to call and so on both of these uh approaches I think are very valid approaches and very interesting approaches and really which one is the most suitable really depends on your use case um and what you're you know trying to achieve um so both are possible in all us and again if you have any interesting insights there or just want to discuss this in general um I'm I'm really happy to uh discuss agent architectures um now I want to wrap on um like uh uh uh very accessible way in which you can sort of uh use Olas which is we have this user friendly no code uh application which is really targeting end users uh to run agents rather than developers building agents it's called Pearl and you can try it out yourself there's currently three agents in there and we have a massive accelerator program at the moment with over a million dollars in um you know uh uh cash and token rewards on top um where if you build um valuable Asian use cases and you think they might be a match go to ol. network accelerator um and you can see the uh program outline um I really would encourage you uh to consider uh that on top of that obviously try out the app and there's y the accelerator a roote for you to bring your own agent in there so at Olas we really have this Vision that we want to bring agents to end users to abstract a lot of the complexities of cryp and we're doing this currently with prediction markets uh social agents and defi agents and there's I'm sure a very wide um extra set of uh agent use cases that could be covered we want to bring them into Pearl so if you're building one of those Agents come talk to us um that's I I think it and um uh just on time so if you want to connect with myself um I'm onx and uh it's davidor enm and obviously here are the sites uh of all us and you can also e m me uh thanks for your time and uh have a nice day