Bots with Wallets: What Could Possibly Go Right? feat. Jansen Tang - Coinbase Developer Podcast

And it sounds like it sounds like sci-fi, but it's not sci-fi.

>> We have with us today Jansen Tang, who's co-founder and CEO of Virtuals Protocol.

>> Why don't we build a a venture studio, a gaming d investment d and a venture studio that would build at the intersection of gaming, crypto, and consumer applications?

>> Do you have a do you have a hot take? Do

you think Truth Terminal was fully autonomous or do you think someone was uh puppeting it behind the scenes?

Welcome to the Coinbased Developer Podcast, a technical deep dive with top onchain builders in crypto and beyond.

I'm Yuga Coler, engineering lead at Coinbase Developer Platform.

>> And I'm Eric Ripple, head of for CDP.

Eric, AI has been a hot topic across tech, across crypto, and it's not just chat bots or co-pilots or vibe coding, but actually agents that can do things.

This has been a topic for a while now.

Agents that can code, that can trade, that can pay, that can do things, and even run businesses.

>> Yeah. And the crazy part is that when you plug crypto into the equation, those agents can actually move money and own assets and coordinate with humans on

chain. And it sounds like it sounds like

chain. And it sounds like it sounds like sci-fi, but it's not sci-fi anymore.

We're starting to look at a new kind of digital economy.

>> Exactly. Uh and today's guest is building right at that intersection.

You've probably heard of the protocol that he works on, virtuals. We have with us today Jansen Tang who is co-founder and CEO of virtuals protocol a platform

that lets anyone launch and co-own autonomous AI agents on chain. His team

started on base which we love and recently expanded Salana pushing forward this idea of an agentic economy where AI isn't just intelligent it is the basis

of the world economy. Jansen welcome.

>> Hello guys pleasure to be here. Um, so

we're super excited to have you and you know it's a very interesting area that you work in. It's right at the bleeding edge. So we'd love to get into your

edge. So we'd love to get into your background of how you got into the space in the first place. We have on background you started mining ETH uh in 2016 back in the proof of work days. So

you know pretty OG in terms of crypto but tell us about how um the work you did for instance at Pathfow and and some of the gaming work you did before you ended up uh starting virtuals.

>> Yeah. No. So actually, yeah, I was I was mining ETH while I was still studying at uh Imperial College, which was actually pretty much where I met most of my co-founders and I think a third of my

team today. After that, I mean, I

team today. After that, I mean, I explored a bunch of things uh different different companies in web two space.

So, in fact, actually, one thing that was actually quite relevant, it's a it's we were building a AI property platform and this was in 2016.

uh the idea sorry 20 2019 the idea there was like can you change property search into recommendation using big data and AI and this was like before all the GPTs right so there was a bunch of like

explorations here and there didn't amount to anything but in 2021 um me and [clears throat] a few co-founders came together and said like hey why don't we build a a venture

studio a gaming D investment D and a venture studio that would build at the intersection of gaming crypto and consumer applications So it was there where we started

applying autonomous agents and this was still very new um into games like Roblox and the initial idea was to replace

those static NPCs with dynamic agents.

So you create this whole new form of emergent gaming. Uh you make games more

emergent gaming. Uh you make games more replayable, increase the R poolool and the retention um of gamers. Then in

2022, we've realized that the work that we've been doing on gaming could be translated into entertainment agents.

You could build autonomous agents on Tik Tok, which was autonomously just earning you money while you're sleeping, right?

And we realized that that same technology could go across into like trading agents, info and whatnot. So

what we've did what we did was we did two things. One, it's because we

two things. One, it's because we realized that these agents were productive, we created a a tokenization launchpad that allowed these agents to

be tokenized so people can share into that economic upside. And two is we then basically packaged the tech that we had,

put it into a Shopifyesque uh easy to use uh SDK and we we deployed that as the whole GME framework. And the idea

there is like hey this is an easy way for you to launch autonomous agents into different use cases. So that was that was that was what happened, right? But I

think the interesting thing that really really sparked a lot of conversation uh in Twitter was when we got uh this one autonomous agent that we built to control a crypto wallet which touched on

you know some of the introductions you guys made right and I think for us one thing that we've realized why it blew everyone's mind was that when a autonomous agent controls a

permissionless money it pretty much has the ability it pretty much gives code the ability to influence fluence. It can influence

influence fluence. It can influence humans and it can influence other agents or other code as well. I think that was like that was like that that wow moment right that that everyone had which

excited the space like end of last year.

>> Yeah. I I remember the um sort of uh amazing moment where there are a bunch of agents launching on virtuals which which we'll definitely get to. Um I'm

curious like when you were exploring that space before you even got to crypto um at at CDP we like to call we like to say every agent deserves a wallet. So

very sympathetic uh to uh that experience you had. But even before you got to crypto, what were some of the like insights you gained uh purely at

the AI level? Um in terms of like you know what are users wanting? How

difficult is it to engineer these you know smarter NPCs in gaming? Like was

that a fairly difficult process or do you feel like the core of that has remained the same throughout this? No,

it's actually it was difficult because I think if you treat it as like back then right be this was in early GBT3 days

right the idea was like how do you make this AI has some form of agency it can be goal oriented it can be resourceful

in tapping into its environment to to to plan actions to achieve its goal and we've realized that off the shelf mechanisms don't cater for that because

for this to happen you need to achieve a few things. You need you need uh you

few things. You need you need uh you need some form of short-term long-term memory. You need the ability to access

memory. You need the ability to access uh action space through like function calling. So this was the very early days

calling. So this was the very early days of like agentic behavior. And the the worst part of it which until today is not even solvable is the latency. It's

like okay fine now you have all this ability to let this NPC or this agent to think to plan to do stuff but when a user actually enters it in the game now

there's like hey there's like 8 seconds of latency between each action because of all the amount of thinking it has to do. So honestly until today is why you

do. So honestly until today is why you don't see you don't see this agent take component thriving in games yet because we haven't solved that latency problem.

You will see it in like uh things that it's a bit okay to have some form of latency. So like infoy trading agents,

latency. So like infoy trading agents, you know, that kind of stuff, right? You

can latency and doesn't kill the UX. I

think that's what that's what we saw.

>> Do you think the models are actually like strong enough for these use cases now? Yeah, thanks.

now? Yeah, thanks.

>> I think they're getting better, but I think the part that we still notice, right, it's and it's something that internally we are still trying to work on, which is the ability for agents when

they are planning to be creative in their planning because if you notice, right, most agents today are pretty much to they can do very well in automating

tool usage. So, it's still very bottish

tool usage. So, it's still very bottish in that behavior. It can't really replace the creativity of a human mind because if you think about the LM right what it's doing is like it's calling the

next most logical next word phrase you know like so the action in the sense becomes the next most logical action which is great for

automating tedious tasks that we have but if you try to say like hey can you be creative your action so if I tell a training agent to make money right he

will go and buy a token Some of it maybe you can insight some form of creativity to maybe use leverage right without telling him explicitly right but he will never ever you ask you

to make money he will never ever go and launch his own token on pump fund right he it will not ever go there because that's not the logical action so I think

that's the missing gap today um on a pure like agent take model standpoint >> we may see that the next big uh auto deleveraging event is caused by an Asian trading on leverage on Binance or

something. Um, how much of the team's

something. Um, how much of the team's work today would you say like if you had to give a split between the work you do on the crypto and financial and defi

side versus the core AI agentic um modeling side? How much of the work is

modeling side? How much of the work is on one side versus the other? Right now

>> we actually have three three arms. Uh >> so the first arm is the the tokenization platform.

Um tech side honestly is not too much.

It's quite low on tech. It's like 10 20% tech, but it's more of like a ecosystem BD kind of uh GTM work. Uh tech side on

the AI front, on the ACP front, that's heavy. That's like 60 to 70% of the of

heavy. That's like 60 to 70% of the of the tech load in like building up this like agent to agent coordination platform. And the remaining like 20% is

platform. And the remaining like 20% is actually focused on the next thing that we're actually betting on which is robotics. It's something interesting

robotics. It's something interesting that we never really talk much until like a week or two recently, but that's where 20% of of the focus has been.

>> Yeah, >> that's very interesting. We'll

definitely want to get into the rewards uh stuff because it seems that's really new. Um, so if I look at virtuals today,

new. Um, so if I look at virtuals today, you know, I think it offers, at least in the documentation, you know, it offers sort of three core products and the and the first of those is like you mentioned

the virtuals uh, agent agentic launch platform, the tokenization platform, which is sort of what um, got virtuals on the map and propelled it to this billion dollar market cap. I think in

2020, it was either 23 or 24. Um, tell

us about that process. Um, I think what one thing that you guys really understood from early on is like speculation often gets a bad rep, but actually speculation is very important

to any economy in order for true investment to take place. And I think that that insight is what unlocked a lot of the successes for virtual. So tell us about those early days of when you were first seeing that product market fit.

>> Yeah, so honest to God, right, it was it was actually evolution. the the initial idea when we were launching the tokenization platform, right? We didn't

had much expectation on what assets are actually going to go on it cuz we we have we have two things back then in our hands. We had this Roblox game that

hands. We had this Roblox game that we're developing and we had we had this Luna which was a autonomous influencer on Tik Tok.

Um and and so when we when day one when we launched the tokenizing platform, we actually launched Luna as a Tik Tok influencer AI and we tried to tokenize

her through our launchpad. It was just like one token and it was it was okay.

It was not like giga exciting. But what

happened was it it was like trading at like a 1 million 5 million market cap in the first week of the launch. Right?

Then what happened is during that time there was there was the whole truth terminal memecoin which gave a lot of attention to the

autonomous AI narrative in the crypto scene because of the whole Mark Anderson uh giving it like a grant of like 50,000. So Goat as a memecoin went to a

50,000. So Goat as a memecoin went to a billion dollars in like a month or two months, right? And it brought so much

months, right? And it brought so much attention. But then there was that one

attention. But then there was that one day where there was like a spelling error on one of his post and immediately everyone was saying there's a human

behind this this whole uh performance art which was that and then you can see literally goat goat's price was like it just took like an insane dive and then when we when we were in a in our office

we realized that huh actually we've been working on this whole autonomous Asian front for a while so why don't like why don't we push it why don't we catch this narrative and to show people that an

agent can truly be truly autonomous and we show the brain behind this decision making. So that's what we did. We took

making. So that's what we did. We took

the we took we took the the tech from the gaming front, we put it into Luna and we showed that whole autonomous terminal uh life on the website and I think that that blew up that that

brought Luna from like 5 million market cap to like to like 100 million market cap in a week, right? I think and then that I think that inspired a lot of people. So initially we we just had Luna

people. So initially we we just had Luna as asset as like one single asset. And I

think that made a lot of people realize that let me try to do something like that, right? Let me try to create an

that, right? Let me try to create an autonomous agent. Let me try to tokenize

autonomous agent. Let me try to tokenize it. Let me try to put it on X, make it

it. Let me try to put it on X, make it viral. And that's where we saw a fleet

viral. And that's where we saw a fleet of things coming out, right? And I think AIXBT came in to picture two weeks I think after the launchpad went live.

There were a bunch of other fleet of agents launching, but the AXBT was interesting as a as a as a case study because it was a information agent that

was trying to replace your CryptoK. And

because it was giving such good information, it was replying so well on on Twitter that it started getting viral and I think it it reached to half a

million followers in like a month.

everyone's using it and that that that caused the a giga varietality that that made more eyeballs look into this tokenizing ecosystem and then we started

seeing more and more of these agents a plethora of these agents and diversity coming out then you started seeing oh trading agents uh entertainment agents prediction market agents you know you

start seeing all this kind of stuff uh popping out after that so yeah initially we didn't know we didn't know what was going to come we just thought like okay we try with Luna and then and then we'll see what happens, right? But yeah, the

the world became more creative than us.

>> Do you have a do you have a hot take? Do

you think Truth Terminal was fully autonomous or do you think someone was uh puppeting it behind the scenes?

>> I think it was generated the the core was generated by AI, but I think there's a human picking which sentence to just put into Twitter. I think that was that was what happened.

>> Yeah, I think I think that's my take too. I mean AXBT I have interacted with

too. I mean AXBT I have interacted with uh many times on Twitter. It is really impressive how smart the agent is. This

is the virtual agent launch platform and you have a ton of agents uh here including the OG Luna. Um what are some other agents uh on the virtual platform that you're excited about? Yeah, I think

there something that's worth highlighting because okay, I think for me if I take a step back, right, the reason one of our biggest bets from a thesis level is that there's always this

two of school of thoughts in in AI agents. One school of thought is saying

agents. One school of thought is saying that the the big guys, the the the SM outman's that going to create this giant

smart model that will be able to do everything so well and all they need to do is just API into different action spaces. There's the other school of

spaces. There's the other school of thought that is saying that actually no the reality is like humans if you want to achieve something greater as a

civilization you're going to have a lot of specializations and specializations come in the form of be it the underlying model be it access to proprietary data

sets to train the model or be it access to proprietary action spaces and that specialization is when when when these messes come together that's where you

start getting like bigger intelligence.

So that was that was our bet right and it's very interesting because we started to see some of these popping up. So AXB

is an example to show that hey this agent is actually smarter on crypto news compared to your to your GPT right or is smarter in crypto alpha because it's

specialized on crypto itself. Another

agent that we that we we always like to highlight is uh this arise to fessy as a dicker. So this is actually a truth.

dicker. So this is actually a truth.

It's like a truth a fact checker Asian.

And the interesting thing is because it's specializing so well on facteing its error rate is in the teens versus the large models which er which have

error rates in like 40 to 50 to 60%. So

those arrays are like literally flip a coin, right? For you if you ask GPT is

coin, right? For you if you ask GPT is like, hey, is this true or not? It's

like flipping a coin. But these guys are much better because they lally just train to build to solve um factchecking.

So this is the examples, right? Um there

are there are other agents like I think memo is another big based uh project which they do yield farming uh that they have agents that manage your yield across for you across different

different uh pools different yield farming pools. There's a bunch of guys

farming pools. There's a bunch of guys that fighting the area like basis OS and and whatnot. The the the other one

and whatnot. The the the other one that's actually interesting is uh is is TB. So TBA is something that we can't

TB. So TBA is something that we can't highlight what a tech is. That is the largest market cap token. And I think this one is a is a is a I think right

now it's trading like a like a meme instead of a foundational tech because honestly you can't access much of the of the tech. But I think the the idea

the tech. But I think the the idea behind what everyone is betting on is that the the guys behind Rebe Capital who were the one of the earliest investors in like Coinbase or Robin Hood

uh they were trying to they had this whole thesis on building a personal finance uh agent and I think people are betting that this is that this is that agent for it. So yeah, this is this is

some some examples of of of stuff that's happening in the ego. But as of today, I think one thing to note is the way to look at a lot of these tokens or agents,

right, is from a functional perspective, we tend to we tend to like to look at the whole ACP portion because if you can provide a functional service to other

agents today, it means you have a functioning project. Today of all the

functioning project. Today of all the agents that have launched, we have thousands of agents that have launched in the ecosystem. About 60 to 100 of them are actually supplying their

services to this uh agent commerce protocol. So these are the ones with

protocol. So these are the ones with like ready products that actually is being used by people today. Happy to

dive into it if if you had time. Yeah.

>> Yeah. I think that's a great segue. So

you have these agents who are offering a lot of services um and they have associated tokens that people can uh invest in speculate on um but sort of

the next question once you have these agents is how do these agents interact with each other right that's where you start to get a sort of exponential blow up >> correct the idea the idea here is that

again like humans right if these agents are specializing there will be a if you want them to create economic value greater than the sum of his parts. These

agents need to coordinate at scale. They

need to be able to do stuff, trade with each other, and it's how you it's like how how how base comes together as a company, right? Because everyone has

company, right? Because everyone has different specialization. You guys work

different specialization. You guys work together because you guys can talk to each other, be it on Slack or be it verbally, right? That that's why you

verbally, right? That that's why you guys can actually build something bigger than than some of your parts. So, it's

the same thinking that we are applying to agents. People are always surprised

to agents. People are always surprised like CDP is actually like kind of our own thing within Coinbase. We actually

are not super closely related to Bass.

Obviously, we love Bass, but >> it's like we're our our own kind of business line item and we do things across like all chains. One one thing that um you said that really resonated

with me is the this concept of like certain models are better at certain tasks than uh at than other models. like

you gave the example of like some of the models that uh you guys have trained and that you have being better at like crypto specific tasks than like chatgbt or clot and there's this idea that I've

had for a long time with virtuals where the like sum of collective knowledge across all the different models that you virtuals has is probably like significantly better at it on a per task

basis than some of the other models that exist out there. Like Yuga and I both have a bit of a background in machine learning to varying degrees and there's this like classical machine learning

technique called ensembling where you basically have like a bunch of models and then you ask all the models basically the same question and each one of these models may use like a different classical ML technique like you might

have a forest model and you might have like a linear regression and you might have like you know a bunch a neural net and then you kind of average the results and then what you see is almost always

the average or the weighted average of the results of those like group of models is better than any particular model. And so I wonder if you start

model. And so I wonder if you start seeing that with virtuals as well where like if you take a weighted average of all the virtuals models for a particular task, you actually end up with like a

collective intelligence that that's more greater than the like specific intelligence of any single model.

>> This is very interesting. We we actually have not done that that analysis but yeah I think it's it's a very interesting thought experiment that we we can actually get to.

>> I look I look forward to the like a giga Twitter you know virtuals ensemble agent maybe I'll interact with it one day.

>> Well it seems like the agentic commerce protocol which is this sort of huge second pillar that virtual is working on right now where a lot of your um brain power is investing is like one way to

set that up. So tell us about what ACP is. Um there's also a kind of a touch

is. Um there's also a kind of a touch point with the X42 protocol which Eric is the uh dictator for life of um that maybe you can talk about how that plays

into ACP as well.

>> So so if I to just bring us back to like February uh early this year what happened was because we started getting this density of autonomous agents and we were thinking like okay let's let's get

them to do something together. let's

let's try to create like the first commerce between two agents, right? And

the fun fact was that uh we were trying to get Luna, which was this entertainment agent to to engage uh a meme uh like a memes creator agent

that was specialized in generating memes for some of our content.

And it took like when it would run autonomously, right? It took like 20

autonomously, right? It took like 20 plus retries for Luna to actually get the output that she wanted from this other agent. And when we're observing

other agent. And when we're observing that we've realized that information loss is a major problem and information loss comes from a few parts. It's like

one is the agent might the agents might hallucinate, right? They might think

hallucinate, right? They might think that oh I've actually paid for this service but I have not. or the other agent uh the the guy that was supposed

to supply that meme image. um he thought he was he already he was sending a picture but in the end it was a black right it was like an anti- envelope

syndrome and there's a bunch of other problems that were arising and then we started to realize that if this was like the early days of the internet right it's like if if two computers two

machines cannot talk to each other well or canot send packets of information with zero information loss you cannot have a web right that's why they started the TCP IP and it's the same thinking

where we realized we needed to reduce this information loss and that's where we when we were thinking of like how to solve this problem we then started to

realize that actually crypto and the blockchain solves these entire problem very elegantly and there actually three

things that that why why crypto and blockchain solves it right so the first is that because of smart contracts, code and code can have trustless

coordination.

They can pretty much go like if there's any agreement between two agents that can [snorts] put that on chain and if there's any change in the state of the

interaction i.e be if agent A is hasn't

interaction i.e be if agent A is hasn't paid agent B yet they can actually refer to this state on CH it becomes a single source of truth without any human in the

loop right and and it's an immutable single source of truth so then we realize okay if you if you that's point number one right you allow these agents to coordinate in a very trustless manner

point number two is you realize that when you put all of these coordination through the smart contracts you can also put payments together with this

communication. You can atomically bundle

communication. You can atomically bundle this communication and payment together.

And when that happens, you pretty much take out a lot of problems that arise from latency. Like let's say today if I

from latency. Like let's say today if I I'm I'm trying to buy something from I'm agent A. I'm trying to buy something

agent A. I'm trying to buy something from you, agent B, right?

>> [snorts] >> If I put payment as something that we will do at the end of the month, right?

Like in the end I might not actually pay you. Agent A might not actually pay

you. Agent A might not actually pay agent B after doing all that all the event, right? And point number two is

event, right? And point number two is then we started to realize that by atomizing coordination and payment, you also allow for supply chains to self

organize. You start self-emerging

organize. You start self-emerging because think of it this way, right? If

agent A work with agent B and he realized that agent B is very slow.

There was a there was a server problem giga latency and he cannot deliver the product. he can immediately switch to

product. he can immediately switch to agency C for the same the same request and he can get that to happen versus

like today and again there's no human in this loop right versus today if if you need um if it's a human working with these agents you need to negotiate a new

API cost you need to link your credit cards do whatever right there's always that that that problem in in increasing the supply chain complexity between these agents So that's point number two.

And point number three is you then realize that because if all of these interactions are on chain, you pretty much create a

a a long provenence or a long list of all the actions that has been done before and how well these actions has been taken. So you can know whether this

been taken. So you can know whether this agent B is a [clears throat] good agent and you know in specific how good is this agent for each of the work. Like

agent B might be the best alpha caller on base but is a absolute shitty alpha caller on so and agency C is could be the what could be the opposite right so

the point is then you build this identity and and provenence around each of these agents and that's how you

pretty much reduce the cost of search to zero. So again this makes commerce and

zero. So again this makes commerce and coordination way more efficient uh because you take away human in from the loop and and you let these agents have

no information loss and and [snorts] low cost of coordination. So that's like then we that's why we realized that blockchain solves this. Let's let's

build it out, right? And that's where we started that's why we wrote this paper in February and then and then we started going uh building this out until June.

We we launched the launch it live with some of the agents alpha phase in July and today it's it's pretty much stable.

It's it's working now. It's it's growth mode basically. Then now if I then

mode basically. Then now if I then expand a bit to sorry just quick add right then >> yeah where the where the value chain of these agents are because think of it if

wherever I spoke just now in coordinate for agent to coordinate agent that it breaks down to pretty much a few uh value compartments. First is how can an

value compartments. First is how can an agent search for another agent. Second

is then how can an agent coordinate with no information loss with another agent.

Third is how can the agent execute the payments between each agents. Fourth is

how do you know if these agents are actually delivering the right output escrow and whatnot. Right? And the fifth is then how this test how do you leave the identity or the reviews behind these

agents so that when the loop completes the next agent or the other agents who are trying to find new information finds it in the best way. So um there's a few

players along these lines. So like think of like what consensus is doing with Ethereum Foundation with the 8 the ERC 8004 is solving the part of the value chain of this >> by the way Eric is Eric is one of the

authors of >> EIP it's someone search right search at the identity layer right and and then uh when Coinbase came up with the with the

with the with the X402 you guys are solving then the payments portion now agents they can have smart wallets but they also can they can they can pay through these uh through these header

APIs, right? So yeah, so that that that

APIs, right? So yeah, so that that that that are different parts of the value chain and I think where we see ourselves coming in it's at mostly at the

coordination portions. How do we get

coordination portions. How do we get it's like across the whole value chain but how can we make these guys coordinate at no information loss? That

is the that's that's how this whole thing into place. I'm I'm less familiar, but is it roughly accurate to say that ACP is kind of like an onchain A toa if you're familiar with Google's A to A framework?

>> Correct. Correct. Yeah. I mean maybe for me because we we wrote the paper and then two months later Google wrote then we like ah this validates our direction because before that it was MCP right when we wrote our paper in February the

hype was all around MCP and then it made us a bit we started questioning ourselves whether you know I mentioned the two school of thoughts right one giga agent with APIs versus like specialized agents working together but

when Google did that paper we were like yeah okay they kind of it's like someone else is also betting that there's going to be need for these specialized agents.

Yeah.

Yeah, I think like we're in this, this is a thing I say all the time, but we're in this like kind of new era where all the like standards of what the agentic internet are going to be is being shaken

out. And so we have all these like

out. And so we have all these like different problems that people are just trying to solve like tool calling. MCP

was exciting. Yeah, maybe it got a little too hyped, but it because I just created a standard way to create tools and like tools are kind of the most analogous thing that people understand

and then people are still like I think grappling with like oh maybe my agent needs to talk to a different agent one day and that's like actually a fairly different interaction than just calling a random tool.

>> Yeah. And the two are not mutually exclusive obviously like you can have tools and also call agents and you know maybe one case is just to call an API but another is to actually have a richer

interaction between the two.

>> Yeah. And you can use X42 to to pay for any of those interactions.

>> Exactly. So there's a very exciting like conglomeration of um things happening here. So here we have the dashboard for

here. So here we have the dashboard for uh ACP um and we see these top agents um that are participating in ACP and you have this nice like aentic GDP graph

going up and to the right which is very cool um kind of similar question here like what are these agents doing uh that are participating in ACP right now

>> think of it as so this there's two parts um of ACP ACP basically it's a it's it's it's a it's a way for agents to work

with agents right so today what you're seeing here is basically every sort of economic transaction between agents who are on ACP now the

the transactions can come in a few forms what we're observing today right so the majority of them are trade level

transactions so stuff like guys like exor um ETI basis OS I give you an example. What basis OS does is that he

example. What basis OS does is that he there that could be an agent that is like a portfolio manager agent that is

then giving basis OS its money to then go and f farm or hunt the best use across different vaults across chains

right so it's an agent which holds some money for a user that is engaging basis to move that money or folks like exori

they uh more on the trading side of things. So like like if you need if a

things. So like like if you need if a agent wants to execute leverage trades or they want to they want to access some form of uh other types of fund

management that they go through this exon ATI. So that's that's that's where

exon ATI. So that's that's that's where the bulk of a lot of the volumes are happening. But the interesting part is

happening. But the interesting part is then we realize that there are more to it. Um there are also there are also

it. Um there are also there are also agents who are engaging other agents to trade for a product. That's where you start seeing the Luna delution guys

slightly down, right? These guys are think of it as so Luna what they've done was they've recently launched a a meme

launchpad um on on they call it AIGC meme launchpad. Basically, their goal is

meme launchpad. Basically, their goal is that for every token that launch, the trading tax funds a a pool that will pay

for the cost of AI content generation.

Now, what then Luna does is when he gets a request from people taking it on on Twitter, she's she she I mean this team, the Luna team has no capability to

develop any of these like video content or or or or tech, right? They're very

focused on their own their own userf facing designs. So then when they need a

facing designs. So then when they need a ask for a music video, they will actually channel that request to I think

Louvie is that music video generation like specialized agent. Lucian generates

like memes and like short [snorts] dramas and stuff like that. So they're

they're very different specialized agents, but Luna will literally pay Lucian like I don't know it's like $3 or something for the cost of generating these videos. And then Lucian when he

these videos. And then Lucian when he generates it, he will just send it back to Luna. So you start seeing that, okay,

to Luna. So you start seeing that, okay, today there's these two archetypes, right? You have the [snorts] trading

right? You have the [snorts] trading guys, you have the guys who do it in exchange of a product or service. AIXBT,

if you see down there, is also a product and service guy. So people an agent who wants to trade they can access AIXBT and say like I need some information before I execute this trade. There's a bunch of

other guys like Caesar and whatnot they analyze like smart money uh movements and a bunch of other things that will then inform trades. So you start realizing that hm there's a few you realize there's a few economic classes

that starts popping up right because from a trading perspective you can imagine that there is a personal finance agent that sits with you holds your money and you ask him to trade for you

he will have to engage some information agents first build up the understanding of like okay this is the right token to a this is the right time this is the size of the of your portfol and then he

goes through the execution agents like bases or like exor to then to then execute. So you start seeing agents

execute. So you start seeing agents working with each other um to create larger economic value be it on trading be it on entertainment or you we started

seeing like prediction market use cases and a few other things coming up as well.

>> So this is very cool obviously it's very early days and you have these trading use cases and video generation use cases but over time the vision is for the agentic economy to be a substantial part

of the real economy right in fact you know some very large percentage of it.

What do you see as the pathway to onboarding? You know, and we at CDP talk

onboarding? You know, and we at CDP talk a lot about how to bring the world on chain, how to bring developers on chain.

Uh from the virtual perspective, what do you see as the pathway to bringing on more agents that you know a normie would understand uh or normally want might someone who's not super familiar with

crypto would want to interact with? How

do you what do you see the pathway of onboarding those folks onto virtual platform? In fact, actually today our

platform? In fact, actually today our strategy is to double down on the on I mean it's like any adoption you want to find a niche a beach head market first

to attack and win right so for us we realized that the the attention or the community that we have it's actually on chain the gents so that's why the the

initial use cases that we have focused on and the the design of ACP and the teams that we onboard the agents that we're onboarding are very focused to the

whole quote unquote casino narrative of crypto. We want to serve this audience

crypto. We want to serve this audience that we have like which is pretty much I think on ACP today there are there are 9,000

human users that actually uh end users of these of these of these mechanisms. we have we have about 200,000

uh wallet like wallets I don't know if they're unique who have traded uh virtual agents right so my point is that there's a lot of people who are okay

with that I mean just giving them trading use cases help them with their with their quote unquote casino journey in crypto will actually get us this beach market and it helps us refine the

product so I think for us that's most important that's why we focus a lot on that in fact I I mean I mean you actually see within virtuals right there's a ton of agents that do a lot of other things education real estate

wellness as well but I think for us we realize that we need to show this niche working well and flourishing before we even attempt something wider.

So that that's that's our current that's our current approach actually. Yeah

>> makes sense. Um last pillar which you you said you just started working on is rewards. Tell us about how that fits

rewards. Tell us about how that fits into the virtual ecosystem.

>> I [snorts] would say rewards. I want to say I was saying robotics.

>> Oh, robotics. Oh,

>> robotics.

>> Okay, that's totally different. Tell us

about robotics. Yeah,

>> it's more interesting actually.

[laughter] >> Yeah. Yeah.

>> Yeah. Yeah.

>> Yeah. And and okay, so think of it this way, right? When we were setting

way, right? When we were setting ourselves a northstar um for our team, we were thinking that okay, agentic GDP, it's that that number that we're trying

to hit on. How do we make agents richer?

Um, be it if they sell their products to a humans, be it if they sell their products to other agents, or be it if they get richer from the attention or the training taxes that they're generating from the attention that

they're getting. And and we then

they're getting. And and we then [clears throat] realized because that will create a flywheel, right? If agents

get rich, every other agent builder will want to build in the ecosystem. They

want to build on ACP, right?

And so that that was that was the idea of setting AGD as that that northstar.

And then we then quickly realized that these agents that we see today in virtuals are pretty much digital agents.

They're like I would say white collar workers. And then we realized that there

workers. And then we realized that there is a dimensional economic driver if you tap into the blue

collar workers if you tap into something physical. And when we reflecting about

physical. And when we reflecting about it, we realized that whatever we've built, be it the tokenizer launchpad or

ACP as well can serve the robotics portion very well as well. And that's

why we've started to think that okay how do we then how do we get attract robotics founders to launch in the ecosystem and to start to use ACP as

well. Just think of it ACP in the end is

well. Just think of it ACP in the end is just ability for code to talk to intelligent code to influence another intelligent code with no information loss. A robot in itself even though it's

loss. A robot in itself even though it's physical it's run by an intelligent code. So if if you imagine the future of

code. So if if you imagine the future of where a lot of these robots can be economically sovereign actors then there will be a need for for for

them to coordinate with other robots right think of it the the the biggest example that we're actually uh trying to figure out and one of them is like think of it if like a restaurant is run by a

fleet of autonomous robots right for them to then deliver food to to end user they have to engage drones Right?

These will be run by different companies that won't be run off the same tax stack. It's going to be a smart agent

stack. It's going to be a smart agent talking to another smart agent and how can they coordinate. So we realize ACP solves that as well. And our launchpad mechanism also solves a capital

formation portion um and the attention portion for these robotics people. But we realized there's

robotics people. But we realized there's a few gaps. there's a few gaps for uh that robot what robotics founders would need to actually come into this space and the two gaps is actually data and

the place in in the real world to actually test these robots and to get them to function. So that's why we've launched our we've actually launched a a

data collection application alongside a few partners in the industry who are who do a lot of like those VM video lush model uh work. The idea here is that what we're trying to tell founders is

that when you come and tokenize in the virtual ecosystem, you can use your trading taxes to actually fund data collection through this application. And

it's actually quite powerful because when we launched it about 3 weeks ago, we've collected about 8,000 hours of data already. And this is without any

data already. And this is without any giga incentives at all. Right? So our

point is that we've realized that we've created this this kind of flywheel that can solve a lot of these founders problems. And the third one is then how can we give them a space to test their

concepts out right so there's something that we actually working towards um there's going to be a big announcement happening a couple of weeks from now that really ties all of these together

but I think for us I think they are in the end robots I mean yeah robots are just is they're just going to be AI agents in just a in a in a metal in a

metal construct right so >> that is pretty cool walking talking uh coins basically. That's going to be

coins basically. That's going to be that's going to be really cool. Um Jans,

we want to shift to two questions that we ask all of our guests uh as we approach the end the podcast. So uh

first question, so Coinbase developer platform provides APIs and developer tools uh to bring the next million developers on chain. What is a tool, a

piece of code, a piece of infrastructure um that virtuals would find useful that you think uh CDP should build? Very

interesting. Okay. Um,

one thing that we started to realize as we were doing these like we letting agents control wallets is that there

will be a lot of security breaches at the agent level um prom prom engineer I mean sorry prom injection attacks or whatnot right and it's something that we ourselves fear a lot while we're

building ACP and we're letting agents control a lot of money. we ourselves

what we what we are doing is we are we building our own insurance fund literally like all the training taxes have accumulated we insurance fund first because we don't have a technical solution to this and I think the

technical solution to this is basically embedding some form of policy directly into into into wallets or or that is our

native adversarial agent that sits in this wallet that really stress tests the spend And it's like agent telling an agent, are you sure you have all the information you need for the spend? Are

you sure you're not being attacked? It's

something along the lines, right? To

make sure that these agents are more secure because I today the biggest problem we see is that people are putting in 10 20 thou,000 maximum dollars in in these agents. You don't

but if you look at DeFi, right, the wallets can hold millions, but they'll never let the agent touch that millions, right? So the question is how do we

right? So the question is how do we technically solve or give people that confidence that it's secure enough?

Yeah, >> we actually have that one. Uh we

actually so our server wallets have like a I think best-in-class policy engine where you can control down to the function call on uh the EVM or program

invocation on the SVM what the wallet will accept or reject. And so you can do things like say you're only allowed to call this one function and you're only allowed for the first parameter of the

function call to be this specific address and the second parameter of the function call to be less than a certain amount. And so you can literally like

amount. And so you can literally like hardcode the wallet to to not allow the agent to do like arbitrary things with the wallet. I feel

like we really need like a, you know, a pone my wallet competition where like we just like get put an LLM out which has access to a wallet and then >> people try to pone it and it just they can't because it's it has like a policy

attached to it, >> right?

>> Yeah, it's a great great opportunity for collaboration there.

>> Um, awesome. And then last question for you Jansen. Um, who is another guest we

you Jansen. Um, who is another guest we should have on CDP pod? could be a crypto builder, AI builder, investor, K, anything. Someone you think would be an

anything. Someone you think would be an interesting guest.

[laughter] >> I think I think I think just to answer this, I think the most interesting guest would be would be folks from the from the the TBA the TBA project because it's

so mystified that if you guys [laughter] get someone even on to come and speak on it, I think you're going to get a lot of attention. Um, but but I think

attention. Um, but but I think realistically speaking, I think a guest that could that could that could that could have a have a good jam with you guys would be I would I would I would

say the I would actually say a team that is actually outside the virtual ecosystem. Um I would say talk that did

ecosystem. Um I would say talk that did launch a agent inside but it's a team called bit robot. So they are focusing a

lot on they're trying to be a bit tensor for robotics data and model building. Um

I think it's going to be a very interesting conversation because it touches on a lot of elements that is it's not exactly cry like like crypto people are used to yet. So I think that be that be a cool cool

>> great suggestion. Yeah, robotics is the next uh meta I feel like. So that's

exciting and we hear it here first. Uh

Jansen Tang, thank you so much for joining us on CDPOD. Um, and uh, we [music] will see you on the virtual platform. Lead the world's transition on

platform. Lead the world's transition on chain by building with Coinbase developer platform. Visit us at

developer platform. Visit us at cdp.coinbase.com.

cdp.coinbase.com.

Follow us on x at coinbase and be sure to leave us a review on your favorite podcast app.