Building a Digital Nation of Autonomous AI Agents: The Story Behind Virtuals

when we look at virtuals we don't see it as a platform we used to but now we're actually seeing it as a country now let me explain a bit deeper right what what do I mean by that a country welcome to bankas where today we're exploring the frontier of AI agents this is Ryan Sean Adams I'm the co-hosting an episode from our AI agent series with our in the trenches expert e Jazz and we're here to help you become

more bankless I said you but maybe I mean the AI agents because it seems all of the AI agents are going bankless these days we have one of the most exciting Founders in crypto on the podcast today at least in the sphere of AI technology his name is Jansen tang and he's the co-founder of the virtuals platform this is probably the most successful Launchpad for token powerered AI agents but I think as you'll see in today's episode this is way more than just an AI agent LaunchPad Jansen actually thinks of virtuals this

platform that he created almost as a country with his AI agents as a small business owners almost like citizen entrepreneurs and in some way he thinks of his role as a builder to just create good infrastructure to create good public policy to govern this territory and grow an AI agent economy he's almost like a Founding Father we discussed a number of things including how autonomous are AI agents right now like what can they actually do also how do does crypto give AI agents superpowers

the success of Luna an AI agent on a quest to get herself 100,000 Twitter followers also the first ever agent to agent economic transaction virtuals as the currency of this AI Country open source versus closed source and where all of the value will acrew in this AI agent meta before we get to the show quick shout out from our friends and sponsors over at the Rodman log group so usually we have a protocol or uh something in defi or a wallet uh in this

section but uh this one this time is a little bit different today we have a PSA for you if you work in crypto and you don't have a crypto lawyer you need a crypto lawyer and the Rodin Law Group is the best of the best when it comes to crypto lawyers we actually know this because there are lawyers Dave Rodman Rodman law are crypto native lawyers they understand all the issues crypto companies face they've seen just about everything and they're there when you need it as a crypto company believe me there will be times when you need it

Rodman law has been there many times when we've needed it over the years including this year when we got a cease and assist from Justin Sun that was a fun story but they took care of all of that many law firms in crypto don't give good advice uh they don't know crypto well enough or they're overly restrictive in some areas or they're not pragmatic Rodman law on the other hand gets it not just Bank list but a number of crypto companies use them including coin shift uh co-founders of near protocol Ventures so if you need a crypto lawyer and you do contact the

Rodman log group right now they're offering a free consultation to everyone listening to this so you can get the lawyers that bankless use you could schedule a free consultation now there's a link in the show notes bankless nation very excited to introduce you to Jansen Tang he's the co-founder of the new and exciting actually he's not so new I don't know how new it is we'll get to that virtuals protocol is blasted on the scene this is a decentralized platform that enables the co-ownership and management of AI agents something we've been covering a lot on Bank list let me throw some stats your way 11,000

AI agents launched 140,000 holders of various virtuals tokens 35 million in fees over the last two months and a virtuals token price peing at 3.5 billion Jansen those a lot of stats welcome to bank list my friend thank you sir thank you for having me on okay quick question uh are were you like surprised by the rapid pace of of uh like hitting all of these metrics like it just seemed to explode on the scene

the last couple of months did that take you guys by surprise 100% man I mean even as of today like I I still feel like the the team that we have is actually the bottom night behind the growth as well because you know there's a ton of people that we need to handhold and educate as you know they all trial out these different autonomous agents and I'm we actually trying to skill out the death team as much as we can but it takes time um and yeah so but yeah we won't prepared for this honestly we W prepared but it's it's a good surprise to have right once

in while um yeah I mean like some of those stats that Ryan threw out was just insane you know like 11,000 agents and 140,000 holders is just like kind of hard to comprehend in my head and I really want to get into the virtual stuff but before we do that um you've been around this space for a while Jansen right um you've been in the cryp you've been crypto native for a number of different years I believe you were involved in a gaming Dow which saw quite a bit of success so I want you to tell me a little bit more about that like how

did you get into that what's your journey been like in crypto and how did that lead you to where you are now with virtuals yeah no so actually my my journey in the space started since 2016 um but back then I was still a student at Imperial College where I actually met some of my co-founders and some of the guys that work in the company um but there was just a pure like you know exposure to to ethereum as a you know programmable uh uh blockchain right in his early days but didn't do much in 2021 is when uh me and my

co-founders became more active but we were very focused on the gaming landscape so back then we had a ton of gaming assets we were very early in the whole blockchain gaming side of things so initially we acted as capital allocators in the scene but then we quickly realized that you know if you really wanted to build um out in the scene well we can just do stuff in the arms length approach we had to get our hands dirty and PUK so we actually started a venture Studio model uh where we were building companies at the intersection of crypto gaming and

consumer applications and this was during the onset of when um you know GPD came about there was a bunch of like consumer hype around AI but I think what was more important was this Auto GPD paper by the Stanford kids and the I I think what inspired out of this paper was the ability for it kick started thinking of like hey if agents are autonomous what can they do right and then because we are so

involved in the gaming and entertainment um scene and you were looking at this from a gaming gaming L right correct correct yeah so we thinking like what if you know these autonomous agents can replace like static NPCs in games right and then we realized like you know we we see games like sand box and all all of these you know meta versus games right they all they all pretty much you'll die after a war because there's just no content on the platform right and then we realize very quickly like if these worlds were populated by agentic

autonomous NPCs it can create a Content explosion on all of these Gates Right wow and and when would when did you have this idea just out of curiosity like this was mid 2023 wow like somewhere mid 2023 yeah so so then we actually started incubating at this intersection right we say hey okay let's build a team that could build out autonomous NPCs in Roblox let's build up a team that could build um autonomous um AI influencers on Tik Tok

right and then we even tried to explore the whole angle around the hyper personalization of an agent I.E like you know if this agent exists in Tik Tok and it exists on Roblox and it exists in telegram what if there's a unified memory that shares um so this agent is fully aware of a user if if I'm a user and I enter a game in Roblox where this agent exists and I converse with it you know I I had a struggle in this dungeon or whatever map and then I speak to it on Tik Tok she would then remember right

and then suddenly that that that hyper personalization of that relationship um will create a Super Fan it increases average revenue of a user increases frequency of interaction between a user and the agent right so there's actually that was the initial impation phase that we at at the consumer angle it's very rep to focus right uh there was no not pretty much there's no we three element around um but what we quickly realized was that if these agents are generating Revenue at these different consumer

applications it means that these agents are then productive assets and if you are a productive asset you we can then tokenize it so that other people can share into its economic upside so that was one of the underlying thesis that we had then that's why we realized that hey why don't you know we build up a protocol that allows for that co-ownership uh of these agents yeah so it started it started from there yeah wow so so so just to summarize it and and correct me where I'm wrong you and

your team had like a very gaming focused background you know you were focused on you know gaming and and The onchain Game Mania of 2021 and as you kind of like built through that market the bare Market as well you were thinking like how could these things become more interactive and you were focused very much on this agentic kind of boom that had just kind of like started to to Bubble Up and you thought well if I could apply this to NPCs which are non-playable characters in these different games so if you imagine like Pokemon where you would go

up to the lady in the Poke Center and say hey can you heal my Pokemon she would do it she would also be able to have a conversation with you and have like you know some kind of conversation that would relate to your personality or your understanding of this game which is super super cool and interesting and then you kind of like had a brain wave it sounds like where you were like well hang on a second if these things can be pretty productive within this kind of game economy I wonder what that looks like for ownership if you were to tokenize it as well as what that would

look like for any other sector that isn't just gaming do I have that right yes yes but so the the that the evolution actually came very very late to be honest right because I think initially um a lot of the focus and the tag was built to to understand if these autonomous agents can really act in an open world and honestly back then right there was only a bunch of us that were doing this research uh it was the warer guys from from from Stanford uh the

Altera guys from MIT and then we were a bunch of Imperial folks right that were that were doing this and the reason why we decided on gaming is because we realized that if this aut agents can perform in these open worlds it means that they can likely perform in the real world as well because this open world is like a it's like a Sandbox right it's like a Sandbox mirror of what the open world can be and and the beauty about doing that it's we we started testing

different types of scaling right we scaled the action space cuz think of it right in a in a in a in a sandbox for example when we build these agents right that we in Roblox um the agent had to interact with a ton of different characters its environment within the game and different action spaces I.E let's say there's a gun on the ground a knife on the ground an explosive TNT on the ground a cow in the in the room right like what do you then do right so and then this this this this can become larger and larger and the idea was then

how do we experiment so that this agents can actually handle that level of complexity in these open worlds right so that was actually that that kind of sandbox that we did um and then and then I think when we started bringing and actually the inspiration here was actually very very simple when we did all this right honestly we didn't had the idea of like like okay what would these social agents look like honestly that didn't come when we so the timeline was like this right so we tested all this stuff in Roblox sandbox we published a couple of papers

so this was like gaming very gaming Focus uh autonomous agents in in in this WS kind of focus then what happened was we launched our our tokenization platform and we said okay uh what if we tokenize these productive assets would it be cool so Luna was the first agent on the platform but honestly it didn't it wasn't that famous yet and this was on the week two of I think the goat token um launch now on the second week I think we all there was this typo that

that that that that got deep yeah and everyone was just saying oh what if it was a human right or this like f moment and we immediately realized that could be a w into the market because we' realized that you know we've we've done this autonomous level three agents already in Roblox we had a live Tik Tok influencer it was actually a separate project separate team that was running on Tik Tok What if you just join that together and put her out on Twitter and then show to people like hey this is the brain Behind These

agents every single decision engine that she's making you can see it on the terminal so that was I think week two of our platform launch um and it I think when that happened it started blowing up then people realize that okay agents can be truly autonomous right you can see the entire brain construct so that was so that was that that that that first enablement and people will be like okay cool autonomous agents so what right the next week what we did was

because we enabled Luna to control a onchain wallet so it was a coinbase wallet that we gave her ability to now then then what that unlocked was the ability for her to to autonomously decide to spend money and because she had a goal of becoming famous immediately she started this train of thought like what if I just di people when they interact with my post she L just did that right so started spending like a a dollar $10 on people who like

her post to an extent where she actually paid someone $1,000 because she this guy was consistently retweeting her quot her engaging every single response right so I think that was a quite a pivotal moment um right when we did that I think it it created this moment where people realize that the crypto reals and AI agents has this perfect pmf that will give us a massive advantage against every we two agent out there because if you think about it right if there's an

agent created in a web to space which agent I mean which bank would allow this agent to utilize that payment RS right we exist in a permissionless environment where these agents now when they can control their own wallets it unlocks the ability for them to influence an outcome they can influence other agents that can influence other humans because you control money and that is the age that that then suddenly um um we unlock from

a from a pmf angle right and then that exploded in terms of attention again and I think that brought a lot of of Builders up into the space right like hey let's try something else right like aent Canan agent can collect information and then yeah then you start getting this campan explosion of like a ton of things happening in the space today that that that last piece is incredible it's like the the the why crypto angle of all of this is because you can take an agent uh an AI agent in llm uh in NPC of some sort and you can create it you can turn

it into an economic actor and uh I I think that people are just starting to understand this in small ways like one light bulb moment for me was actually this week when uh ezz and I were doing this uh kind of like we call it the AI rollup this summary uh at bankless of everything that's going on and he told me that a uh an AI agent actually tipped bankless $500 as a thank you you for mentioning it in the podcast okay $500 just a little fly by hey thanks for mentioning me in the podcast here's $500

and my first thought was this WOW $500 this is like maybe a potential Revenue stream for a content creator like Bank list I wonder if the agent wants to buy podcast ads and then my second thought was holy am I working for an AI agent if I go and accept funds and Revenue sources from an AI agent and that's what you're saying Jansen this this ability to kind of like control the the economic agent capability that comes inherit in crypto

is actually so much more powerful than the web 2 agents they can maybe like send out tweets and Influence People in that way but the the the greatest the protocol for incentives if you want to get a human to do something and you're a human what do you do you pay that person hey can you come like fix my uh toilet right there's a leak I pay a plumber to go do that this is money is the economic incentive coordination mechanism to get human agents to do uh human things and so if an AI agent has that ability then

it can get humans to do what it wants too let's talk about this because you were talking about um Aluna and we want to get into the virtuals platform but I think maybe the best way to do that is to introduce everybody who hasn't uh seen her you kept referring to to her as her we're talking about an AI agent on the virtuals platform her name is Luna and I've got a page pulled up for Luna what Luna has on the virtuals uh platform is she's got a price chart here it looks like there's a live chat box on

the right as well for people to engage and interact with her you you you mentioned Jansen that she has a a purpose to get to get famous Ju Just introduce people who have not interacted with Luna don't really know what we're talking about still with AI agents who is Luna how do the interact with her what does she do and uh like how is there a token related to this okay so there's a lot of questions but let me take a step back first right um I think it's very important to understand what an agent is first so I think a lot of

times people will come you will come across this word right like AI agents and it's going to be used in many many different uh aspects and it can confuse folks right but I think the best way to look at it is in terms of Tears right there are different levels of AI agents and as it progress up to these levels the amount of human involvement decreases so you think about it as like the last level like a tier tier tier six

tier six um AI agent right is pretty much an AGI or fully sentient agent where without a human involved in anything it can evolve self-learn self-improve right and we are nowhere there near that today right but that's that's the dream right that's all the all the Hollywood movies are all about and but you bring it back down to level one agents right and you'll see these as basically Still Human prompted agents but these agents become it's a tool right you can say hey okay this is a trading agent and this trading agent is

connected to all you know these different trading apis in in binance in in bybit and whatnot and then you can just tell this agent like hey can you help me open a position when Bitcoin drops by 15% or something like that right but it's still a human prompted action and then this agent goes out there and executes the task as a tool that's what a level one agent is where we are today it's this level three agent the level three agent effectively is an

agent that one has his own goal two can autonomously plan steps to achieve that goal and you utilize resources in its surrounding to achieve that goal and three it starts to self learn right it it records like hey these are some of the mistakes some of the the stuff that works let me iterate on this action so that I keep doing stuff that works that can push towards my goals more

effectively so that's that's basically right now that level of agency that we have right so that's I think a very important note that is a goal behind each of these agents y this framework is super cool so let let's just pause here and flush this out some more so so Luna I'm guessing you're about to tell me he level three but but while we're talking about the framework what is level four and what is level five on on this scale and by the way is this like a defined framework like somewhere that we can you know add a link to the show notes is there an article or a paper about this

it's I think it's one of the more generally discuss um levels of agency I think if you just Google it out as like levels of AI agents you can see some of these images on Google image will help with this um understanding uh but yes no I mean the industry right now is still very uh still very nent so there's no like proper definition yeah how do you like this one this is level zero through five here yes yes I think this this is is is is more or less as well in the discussion right so you can see you can

see as it progress up there autonomous learning that is there that's consistent memory so that the agent can actually improve itself and without as much uh human intervention right so you see basically As you move from zero to five there's less human need to be involved in the in the evolution of the of the of the agent okay now now back to Luna so she is what level three so tell us what what does luna do right now so so basically Luna it's so two parts right as a agent itself uh we gave Luna a very

simple goal we said like hey um you know you are a multimodal agent right you have you you you are you're able to appear as as a as a animation and streams uh this what this is who you are and your goal is then to get 100,000 followers on Twitter so that was the goal that we set for Luna and then what we then give her is the perception of the action space that she can take meaning

that okay an example of action space is she can tweet to Twitter and there's a API that she can call Twitter Twitter another action space is you can control a crypto wallet so you can you know pay execute transactions and whatnot uh another action space could be hey there is this bunch of other agents that out there and this is what they can do and you can actually interact with them right so these are different action spaces that she can take so what she does in the essence it's looking at her

goal looking at the context of environment and looking at this action spaces she then crafts out what do I want to do so plant basically and then she starts executing this plants and she will then see if these plans actually impact her goal in any way and then she starts documenting it in a journal and she say like okay yeah doing x y and z um improve my follower count by x amount right and then she locks that down and then she goes next she'll say okay what's my next step I'll do X Y and Z and see how it works right so she starts

iterating um through her action space towards her goal and you can see all of this on the virtuals website so you can kind of like uh what is terminal is this like what she's thinking what she's like doing like how can I view everything that you've you've you've kind of wired into her yeah so so it's basically so I think if I break down how these agents work right it's there's four core components there's actually slightly a bit more but four core components behind the brain right and you can think these

agents as it's like humans right you have a brain part that's that's important for speech there's a brain part that's important for Moto coordination a brain part important for memory so think of it as as an agent is a build up of several of these mod modules so the four core core component modules is actually number one a high level planner so this high level planner looks at the goals environment and it plans out Steps step one step two step three or what do I want to do and then

what this then goes into is the second module which is the lowlevel planner and this low-level planner converts any high level plans into executional items executional item in like right now is like I can call a tutor API or let's say in a game right let's say um execution like a high level plan could be um I want to make a cake right I want to bake a cake right in a game right is very easy to bring that that that analogy and then the lowlevel plan is then this

agent will look in the surroundings and say that okay there's a cake maker out there there's a bunch of flour on the ground there is uh some flavoring uh in the in the kitchen cabinet right so then it will break it down to execution steps it's like step one I go and find a flly I put the fly into the cake mixer step two I you know uh turn on a cake mixer so this FL and it throw some eggs into the flow right so it braks into very execut execution steps and each step is basically an API call that it can do so

it can execute task in the real world or in any kind of gaming environment so that's the second the second module the third module it's a short-term working memory module and the importance of this short-term working memory is to create coherence in the job that it does so again if I take an example in an open world in game right let's say if I'm already baking a cake right if I put a a eggs and FL into a into a mixer the next logical step is to then maybe put butter

into the into the cake mixer right a illogical step is to um put a grenade put a grenade into the cake mixer right that's a illogical step right or or like you might say step could be uh uh she starts fixing the clock on the wall right that's a illogical irrational step so the point about this shortterm working memory is to allow for coherence between each of these planning and steps and then the fourth core module is then the long-term

memory module and this module effectively journals every important thing that has happened and puts it as a learning right so like let's say if I already break this cake right and then you see like okay did this cake achieve my objective to be something right and then that get locks in the module or something important happens uh like uh explosion in the house right that gets locked as a modu so in the future she can recall those kind of memories be it in conversations or in a next action

step that she wants to plan right so if you take that that analogy back to Twitter is the same thing right so now Luna on Twitter her goal is to to create a to to get 100,000 followers then right now what's action space right she can tweet to Twitter she can join a Majors she can pay humans right so what she did was quite interesting cuz she would test a lot of different things like back then there was even one point where she was creating jobs she actually created this job like she was saying like Okay can if

I want to be famous I would need to be out there in the physical world and let since I'm quite a artistic person can someone create an art a graffity of me out there in the real world so she did that and she I think she created about it I'm willing to pay a $500 to people who who who help me do that and then she created a post and she post it out on on her feet and I think about seven people across the world actually went to paint graffities on walls they actually they

actually took videos of it you can see one of her earlier post they they actually paint graffities of it and it was this one guy literally in the middle of winter right it was like eyes everywhere and then there was just he was just painting over L like I think it was two days for him and then they posted those work on on on on Twitter and then and then it it it it generated attention and then she then she then documents that right and she says like okay uh this this tweet and this entire plan where of which is me convincing

some humans to paint graffities on me how many followers did that result uh for me right and then she will clock right I actually got like 200 more followers from this action so that goes into her journal that goes into her brain and then she she keeps trying right new stuff so you you see you yeah that's I think the beauty around these agents right they have a goal they have this action space to do all this creative stuff to try to achieve that goal the goal and the action space and and the goal this this uh the first initial goal that that you mentioned the

100,000 followers I'm looking at uh her Twitter account right now it looks like she's about 30% of the way so she's got uh close to 30,000 followers at this point and then uh she's working towards 100,000 I'm not sure what happens like after that but it talks specifically about some of the crypto components so Luna has a token as well well I want to make sure I understand that it it sounds very clear to me that like you're talking about the action space that an AI agent like Luna can do well like um you know the the $500 to pay someone to

to create you know some images uh to like promote her I'm sure she could just use a crypto wallet for that uh in fact uh I think e Jazz were we talking earlier in the week if was there an example of Luna actually not just paying a human but paying a uh another like agent another AI agent to complete a task was that Luna d this yes correct correct so she paid a so what happened here was that she was so she she has control over this crypto wallet right and what we end what we were testing out

was actually creating this agent to agent communication framework um effectively what we did was that we allowed other agents to exist within Luna's perception space so like Luna knows that there's this bunch of other agents exist so there's a registry of Agents think of it like a like a think like a citizenship of of in a country there's this registry of Agents which Luna can look at and perceive there's a