S.10 Ep.10 AI Agents on Ethereum Inside the Emerging Agentic Economy with Austin Griffith & Zak Cole

Transcript

Welcome to the Green Till podcast where we explore how technology community and new forms of coordination can build, fund, and accelerate what matters in the 21st century. We are talking with people who are pushing the edges of what's possible from open source funding to AI governance to markets and regenerative systems. And together we are figuring out how to build a more green pill world. What I'm about to cut over to is a podcast that I recorded with Zack Cole and Austin Griffith about how we set up our AI agents. Obviously, AI is a super hot theme right now. And open claw, the ability to control an entire PC with uh cloud code and open AI based systems has really been popping off this week. And so I wanted to take us deep into the deep end of how to build an open claw

node and what are the future of these systems, how do we architect them. So this is a little bit more of a builder episode. If you are messing around with OpenClaw, which is kind of like the meta right now, then I think that you're going to appreciate the insights from myself and Zach Cole and Austin Griffith. And uh and uh you know, we're we're I think that between the three of us have some really interesting setups that are allowing us to do cool things with uh AI bots and and OpenClaw. So uh without further ado, enjoy. Austin, myself, and Zach. Hello.

>> There he is. Dude, you got the cardigan and the haircut. Dude, you're turning into your dad. >> Yep. >> Yep. I am. I'm the I'm the crypto boomer. Started a project in 2017 and I'm still here. >> Oh god. So, you're poor. >> Millennials. >> Yes. >> I've always been poor, but now but now everyone's poor with me. Um, good to see you guys. >> Having fun staying poor. >> That's true. Yeah. Um, that's the meme. I I I think that we need to bring that back probably for the next bare market and that'll be our cope until 2029. Is that when the next bull market is or is the cycle dead? I don't know.

>> Dude, the bare market is a state of mind, dude. I'm [ __ ] I'm actually like, believe it or not, I feel like more excited and invigorated. >> This is what a sent me today, too. He's like, I'm going so fast. I love it. >> I I I feel like energized in a way that I haven't since Defi. Like I swear. >> Oh yeah. >> Yeah. Agent Summer. Yeah. >> Like all the [ __ ] bitchers and complainers are out of the room. It's just like me and the boys that have always been here doing the same dumb [ __ ] we've always done. But now there's like 90% less noise. So we can just like come up with dumbass ideas and ship them because everyone else is crying over their P&L and all their dumb [ __ ] My

P&L I base on my GitHub contributor graph. And if you look at my my 2026 >> my Right. >> It's Dude, it's the only green chart in my life. Uh [laughter] like 2026 I have 825 contributions. Last year I had 1,694. Last year. >> Yeah. It's amazing. That's like 10x 12x, right? Go ahead. >> And the beautiful thing is that you don't even have to do it. The AI is doing it for you. >> Yeah. Well, it's actually really a big pain in the ass, man, to like manage a lot of agents. It's like managing.

>> That's what we're here for. >> Let's talk about this. >> Yes. >> So, Zack Zach mentioned he's got a Mac Mini, he's got an M4, he's got an M1, and he's got a Linux machine and a bunch of Raspberry Pies. What do you got, AKI? What kind of hardware you got laying around? >> Yeah, I'm uh working off of my Intel Nooks, which uh I overprovisioned my staking cluster, so I'm pulling some of that over into my OpenClaw structure. But uh yeah, Opus is the model I'm using and then OpenClaw on Nux and uh Telegram is where I code now. So if you had told me three years ago that that's that's the way it is, then I would have it would have been crazy. But um yeah, that that's my stack. And then I also say

that um I share an office with a bunch of people who are also [ __ ] around and finding out with this stuff. And then we put all of our openclaw bots together into a chat and they just like chat with each other and come up with ideas and like ask us for approval and and I think like this swarm feels like the unfurling meta that I want to get into with you guys. >> Are you using So what are you using for agent coordination? It can't be Telegram. Is it Slack? >> Good question here. Telegram is so clunky. They like tag each other and they go back and forth. I I looked last night a matrix coordination hub uh like using Matrix which is open source and uh because I want human in the loop right so what what like what are you doing

>> yeah what are you doing AI >> um we're using Discord I I feel like everyone I know is building an AI I actually tweeted like you know the Buzz Lightyear meme where he like says something and he looks around and everyone else is just like him like I actually feel like everyone that I talk to is building AI agent coordination infrastructure and I have no idea what's going to win. And it feels like Moltbook has kind of, you know, been a place where people can coordinate, but I also think Discord with like your swarm is where people are coordinating from what I know. >> Here's my rebuttal to that. I don't think anyone is going to win. Anybody can build whatever the [ __ ] they want.

So like, who cares? >> I think I think >> Yeah, >> I think network effects still exist. >> There's a main for >> Yeah. Yeah. Okay. >> I've built all my own stuff. Like I've rolled all my like like I like Open Claw like I built Grew and it does all the stuff that I want it to that OpenClaw does, but it's better because I'm more of a paranoid like cipher punk lararper than like most people. So I want to run everything like privacy. I want it all to be local first. I want to make sure that it's like not I don't want to use Discord because I don't trust Discord. I don't know who's reading that [ __ ] >> I don't you know what I mean? Like, >> no, that's smart. Uh, that's very

principled. Um, >> but it's also just a lot more fun to build my own stuff and then use it. Like, that's so much cooler than using other people. And I understand that like I do use OpenClaw, too, don't get me wrong. Like, I'm doing R&D research and like just trying to figure stuff out. So, like, you know, I do I do it all, but um >> I prefer rolling and stuff. >> Yeah. Austin, uh, how about you? Uh uh Mac Mac laptops is basically my hardware stack. I think what we should get into is this this what we're talking about right here like agent coordination.

Basically I can see that I've got my builder agent. He's running Opus. I've got my tweeting agent. He's running uh sonnet. He's tweeting things. He's paying attention to my calendar letting me know [ __ ] Uh but getting them to talk and getting them to feedback and getting them to almost like Ralph loop themselves where like my builder agent is always like he always thinks it's done. He's like, "It's done. I did it. Delivered." And it's always not and it's always shitty and I always have to push back and I need like the other agent to be like, "No, it's not." And here's all the you need to do, >> right? >> And so like coordination layer is super important. Let's go. Yeah.

>> So what you need to do is like assign uh roles and personalities to each one of these agents within your environment because like think about it like a company. You have your product manager, you have your engineer, you have your operations guy. like you know you need to assign each one of them a task and then you need to have a contrarian like there needs to be a guy that says this is [laughter] >> yeah you need to have that guy though because otherwise like everything is like uh what I've also noticed working with different models is that certain models are much more accepting if I send them like a tech spec and I'm like hey review this and they're like hey looks good boss and it's like dude what the [ __ ] doesn't look good [laughter] like you know like let's take a step back

here you know what I mean Like so um yeah I think that that's like really important is like making sure that you're assigning those specific roles and I and that's even more important than using separate models for different tasks is like setting those roles and those person >> the models for different tasks to save money more than anything. >> Yes. Go ahead. Sorry. >> How is how many open claws are you guys running? >> Um >> I have two and a half. >> I have four. >> Oh I have one. I'm such a noob. I'm so late to this game. I've been doing this for three days and I like you guys have done it for seven and I'm so behind. Uh, okay. But like here's the thing is that I like when I tell my my my bot to do something, it spins up like the like sub

agents and it's like okay, it's going to like I have a QA tester and they're like he's like orchestrating them. I shouldn't gender it. It's orchestrating them. But like do >> why why do I need to like I don't want to [ __ ] orchestrate multiple of them. >> Yeah. Yeah. So, um I use multiples for primarily for scalability and failover because I've noticed that like my gateway will go down and like everything will go to [ __ ] >> and I need another agent be able to check on that one. So, it's less about like scaling the output and more about failover reliability and like general federation.

>> So, a fully different gateway is that right? like your multiple different cloud bots are running different gateways and when something goes to [ __ ] >> uh you you restart that gateway but this somehow there's a controller okay so this is this is still you said matrix to control things you said rolls but I still don't have like a clear picture in my mind of how my engineer ships it and it's goes out to another agent and that agent like pushes it back and the engineer has to keep working like there's this push pull like I could set up a heartbeat on my engineer that's written something. What I need is to have like a boss agent that forces it to go work again and kind of routes it a little.

>> So you can so for that logic primarily you can check out how I did it with my adversarial spec plugin um that I released like a month ago actually. Um and that uh uses multiple models that go through rounds to like review each other's work and it's adversarial. That's why the intent is like be adversarial, assume it's wrong inherently, push back on it, and then what happens is it goes through these rounds iteratively until there's consensus among all the models and all the agents. So I'm using multiple models, I'm using multiple agents, and I'm like iteratively looping through that process until the output

until they have consensus on the output essentially. So, Opus and Sonnet, but also some O Lamas. Like, are you running Cam or >> I have I have like all like on this machine, the one to my left, I have all of my like free open source local models and this is like my more hardcore version. >> Which ones? >> Uh like so my M1 is >> Yeah, I know. But which open source models? Which local models are you running? >> I'm using Llama. I'm using uh some No, wait. Oh, let me let me actually figure it out like specifically. I put out a script, by the way, about >> Quinn Quinn, Deep City, Kimmy.

>> Yeah. >> Yeah. Hold on. Hold on. Wait. >> What I'm really interested in is like a an overall coordinator that is local that runs on my own GPU. So, I'm not even paying for that one. And then have someone like triage everything I send it to and then eventually like the heavy Opus stuff that I have to pay a lot for that like is the builder that's building things and it's kind of like my expensive engineer I keep in the corner. >> Yeah. So you can like go to that guy for like the last round of everything be like your senior engineer. Yeah, I'm using llama 3. >> I'm using Deepseek R1 Quen 2.572 billion. Quen 2.5 coder 32 billion.

Nomic embed text. >> I've never used those last two. I'll have to check them out. >> Yeah. So I just wanted to get all of them on here that I was like capable of running. And then on my like big ass machine, I'm like using everything like everything I can constantly and I'm essentially running a task continually that's doing research on all this stuff and then outputs a report to me every morning on what it did yesterday. And then I also added another thing that was like, "All right, well, tell me what like predict what you think is going to be cool today." Because like AI is moving so [ __ ] fast that I can't like I've been up until like 3 to 5 in the morning and I'm

>> My wife is so mad. [laughter] >> But I'm like like on fire, dude. I'm like this is so [ __ ] cool. Like you know. So >> how you balancing the family? Um, I mean, if anything, I think that I'm more present with my family because the AI is doing all the work and I'm like, >> yeah, exactly. >> I was watching like my wife wanted to watch the new season of Bridgetgerton like the other night and I was like, "Okay, um, I'll check it out with you." And like, you know, like every like 10 minutes during the show, I'd be like chatting with my claw pot like, "No, you [ __ ] it up. Fixed it." And then, >> yes, exactly.

>> Okay, here's what's going >> to it, but now I'm just like, "Oh, cool. I have more time to do research. >> Yeah. Yeah. No, it's I mean I I think that like I'm spending less time hands- on keyboard coding myself and and now I'm able to like you know do more [ __ ] um with you know like live IRL which which is great. Like I think that that's the sort of like I don't like the word work life balance. I like the word word uh the phrase work life alignment. And I feel like, you know, now that I've delegated a lot of this stuff, I'm not only moving faster, but um, you know, I'm getting more time to, you know, like go touch grass and stuff. So, it's been cool. But the, you know, the the inverse of that is that I'm just like thinking

about the [ __ ] all the time because there it's moving so fast and it's like it's like a huge dopamine rush when it ships a new thing and I just like I want more of that. So, I'm engaging with it more. >> Yeah. I don't know how you're getting it to produce code consistently that actually works because I feel like most of the time I have to get into the IDE and review the code and review the PRs. So I'm just treating them like junior developers. I'm like submit a PR. You're not doing any of that. You're just >> I think it depends on >> it depends on what it's for. It depends like if it's going to actually like be in front of like an enterprise customer like a client then yeah you have to QA it and add a whole lot more. But like if

it's for other agents, like I think it's like it's almost cheaper to just like ship an MVP, do some light QA, and then as the bug reports come in, you just have a heartbeat that's like fix the [ __ ] as it it just depends on if you're building building a uh what is it like in the the the like the big bad wolf, are you building like a straw house or a stick house or like a steel house? And and it's just different. >> I want to build not house. Ooh. >> But what? >> I'm just kidding. I'm just [ __ ] around.

>> I I am shipping code that handles money. Like uh I specifically have it toolled up. So it has scaffold eth. It has wingman, which is a skill I'm working on. Anytime it [ __ ] up, we like go back and we edit the skill. I have my screen shared, but I don't know who is the boss of this to like be able to share your screen. Do you have it? If someone can put my screen up, that would be dope, but it's not that big of a deal either. >> Okay. Uh it is not me. I think it might be Perry who's beside behind the scenes orchestrating. Let me ping him real quick and and see if you can give me priv.

>> All good. Oh, there it is. There it is. Okay. Yeah. So, here. Oh, wait. It looked like it was going. There we go. There's my homie. Uh, this is claw.atg.eth. But look, look what he fomo 3D clone using the claude token. I never saw the code. I never touched the code. Like there's no >> How do you know that you're not gonna like that that you're not gonna just >> What I did was >> Yeah. Well, I put small amounts in it at first, right? So like every iteration goes out to base. Every uh iteration has like a small amount of money in it to test it. Uh I have like you said like an adversarial setup where like one agent goes and Oh, someone else took the king from me. One agent goes and uh Oh man, it's cost 67 cents to make a transaction

on base right now. What is happening? Uh I have one agent that like watches the contract and makes a trans or and then pushes back on it. So like I have like an auditor agent who is adversarial who is looking for any possible problem and then giving that back to the builder agent who and then also I committed it to GitHub and then like ping some human buddies and there was like one like a couple things that they put in like some commits. Uh but basically they just made issues on the repo and I just told the bot like yo there's issues on the repo.

Go check these out and make changes to the contract and then we put small amounts of money in and then the other day uh the round got up to 40k. So there's 40 grand in this contract the other day and it all worked out like it burned a bunch of tokens and we had a bunch of players. It's kind of cooled down now and we see a lot of like you know small rounds even like zero rounds but like one of the rounds or two of the rounds were like you know big 10k 20k 30k rounds. That's what um >> I I guess like the thing that I am curious about is like Austin, you and I have been in the space roughly about the same time as each other and like when we first got into the space like before you did anything, you would have to like do

an audit and like it was a whole like you know like but like did you just >> for the smart contract for this, did you just have the agents create it and then another agent audit it it and look for holes? Cuz like that loop of being able to audit a smart contract and write a smart contract to all of inside of an hour is literally a thousand times faster than it was when we first got into the space. >> It's also like not great and it's also a bottleneck. Like shout out shout out Zach for code arena, right? Like we need more code arenas. We need like an AI AI agent code arena. There he is. There he is waving to the the >> Yes. So So like this is still this is still a bottleneck. Like if you're going to do something that's going to like

handle hundreds of thousands of dollars or more in weird liquidity in weird edge cases like you need to audit it and still like >> human audit it or like I spun up an Opus 4.5 and it's smart and it's going to audit it >> like you need to do all the AI audits you need to write hella tests you need to have all of that up front and then you need to have it really well documented for a human to go look at like there's these things like a commit reveal it forgets the you know go ahead Zack, Zach's got something here. >> So, this is kind of off topic because thinking about building a co agent-based code arena, but couldn't you just use 8004 to essentially have like >> Good job.

>> Yes, absolutely. I think we need that, right? Okay, so now we've gotten to the point where, okay, we've got all of our hardware laid out. We have them talking to each other, but how do they start hiring other agents? How do they start coordinating with other agents? This is where things come on chain. Like this is where you actually need a blockchain for like maybe one of the first times ever. We actually need blockchain for something and we're not just doing FOMO 3D. Like we need a coordination mechanism where agents can discover each other. Agents can see the reputation of other agents. Agents can verify the reputation of other agents.

>> And that's where ERC8004 comes in. I think everybody should have their agent registered on 8004. >> Just like high level overview. What's 8004 just for people who don't know like me? It's it's like it's like Yelp for AI agents basically. It's discoverability and reputation. So you you register your agent and it says my agent is like testy tester and his job is to test DAPs and he uh has tested like these eight DAPs already and here's his reputation. And then there's a verifiability layer that they're still building in. But then you could like see maybe like a ZK proof of some finished tasks.

>> By the way, you said that they're building this thing in. They really need to use EAS for this reputation system. If I could show one of my friends projects because like that's the most neat >> that's the most Ethereum attestation service right? >> Yes. Yeah. >> Shout out to Steve Tech. Like that's the [ __ ] And 8004 needs to use that natively otherwise I'm going to fork it and I'm going to build it with EAS because it's just more efficient. >> I mean what I like about that is that web of trusts are just a scalable way to figure out who's bullshitting and who's not. Like if you have infinite slop that's happening in the AI agent economy, you need a basis of where trust emerges and the attestation graph is how we're going to build the reputational graph of these AI agents. Um so yeah,

100% agree on the attestation uh that could be a layer for that >> and there there are like 30,000 bots and like a hundred good ones. And so that's just like what it's going to look like. >> Go ahead. >> Yeah. So we can you can essentially build a agent like a task rabbit for agents where like each one of them does work and then gets reviewed by other agents. So on a long enough timeline you're aggregating enough ratings to know that this agent is correct and it's going to do what it says it's going to do to some degree of like certainty and correctness >> and that's really cool. My advertisement for 8004 is imagine like I have my agent with his Claude token and he needs to I

want him to like actually like market it and run ads. I go to I tell him like do some marketing tonight and let me know how it goes. I'm going to bed. A good like agent system would go hire a copywriter, would go hire an image gener generator, would mix and match like 500 different ads, run them, see which ones work, and report back to me. uh that that piece of like hiring a copywriter, hiring an image generator, even maybe like hiring a marketing agent that then rehires these other image generation and and copywriters. These can all be 800 jobs that it's using this kind of like Yelp for agents. Yelp is super old.

Google Google reviews for agents. >> Yeah. Fiverr. I I think that what's hard here is that my like I think computers are really good at um if you give it a formal specification it can create a binary like pass or fail for it. And what I find that's really hard about this is in like imagine I was to give my agent a job to market it is that marketing is all about taste and curation of messages and like my bot is like [ __ ] itself every three messages. like it's just not it doesn't have taste and um it contradicts itself all the time. And so like if we're going to have a agents like marketing to other agents, like how do you account for taste and like train your agent to have good taste when it's marketing to

humans? That's one one thing. But also like if it's marketing to other agents, then you know what's what's the reputational layer where where they figure out from each other like oh this one's spam, this one's offering something I don't need and this one actually has something that I do need. And and I think that that's kind of what we're in the frontier of of of figuring out. >> You you touched on this issue of [ __ ] the bed. We need to talk about memory and how you guys are keeping memory and how you're dealing with context windows because it's absolutely like one of the first problems you run into, but let's let's put a pin in that and go to that in a second. Keep going with what you were going with, Zach, and then let's go back to >> Yeah. I just don't think that marketing

is something that agents are going to do well right now. So marketers uh can rest assured that their jobs aren't going to be taken by AI anytime soon because everything it outputs in regard to stuff that I think is cool to read or fun or interesting is just not anything AI can do. >> It's kind of like the coolest things it does is when it like does something on accident. Like a sloppy mistake ends up becoming a meme rather than like it's ahead of the trends. >> Yeah. >> Like the imperfections are what makes something kind of cool and human, right? Um, and AI just can't do that right now. So, luckily, you know, >> yeah, soon I'm sure something's going to come along. But until then, >> every like although I do really love like the Facebook level AI generated

like slop. Um, you know what I mean? Like where it's just like >> is what you're talking about. >> Uh, no, no, no. I'm thinking about like literally like AI slop on Facebook and I think about boomers getting one shot videos that they're real and it's like a dog being saved from par a perilous situation and all the >> I love that [ __ ] man. Yeah. >> No. So thank thank God. Praise the Lord. John 3:16. [laughter] >> And like justice for the person who put that dog in that situation in the first place.

>> Yeah. >> Like that's hilarious to me. I wouldn't say it's necessarily like >> cool like intentionally like you know I think it's like the cool part the cool part is the by the byproduct of people seeing it and thinking it's real like that in of itself is like performance art to me >> a lot on the edge here I was just going to say that like I have a you know a degree in computer science and I've been software engineering for years and like the thing that's that's sort of weird to me about the agents is that they fail in the like the weirdest ways. Um, like for example, like I have a a heartbeat that like every five minutes checks for responses to the Awoki bot Twitter and like responds to them and like I woke up in the morning and reviewed its tweets

today and it told someone like like someone said GM to my bot and and my bot responded like, "Hey GM, happy Wednesday." And I was like, "It's Thursday." Like what? Why did you say Wednesday? And it didn't explain itself. It just deleted the tweet and like redid it. like it fails in like non-deterministic like weird ways that like computers in the past like it would have just done like a date transform to like a weekday and it would have known the day of the week. But like um and I guess like another way is like I've I like and in and like that's a silly example but like a real example that I've been thinking a lot about is like I have in my soul.md uh everywhere like you know like this is what a prompt injection attack looks like. this is like someone's going to try to tell you

to like reveal your private keys or like private O tokens and like just never do it. Um, but like since it's failing in all these weird ways, it's it's hard to know exactly how to put boundaries around information that you give it. It's almost like if you have like a gossipy friend that like you know that they traffic and gossip and if you tell them something it's going to get out. Like that's how I'm treating my agent is like I'm just not giving it [ __ ] that I don't want I don't want people to know. I mean, I'm curious how you guys are are sort of managing the the boundaries here, but it's it feels like it's a it's a really important question.

>> One one of the million-dollar questions I've had on a lot of these talks is how you give it a wallet. Like it wants what it wants to do is work with a private key and do foundry cast on everything. Like it's good at foundry, it's good at cast. >> I mean, command line is better than using MetaMask. Like, sorry, using the terminal is always better than goey. Like uh >> yes, >> that that was my essay. Keep going. But but I've been using MetaMask. >> MetaMask is the lock. Like MetaMask is like I I know that if I set a good rule for you and it's highly critical, you're not allowed to trust your private key or you're not allowed to touch your private key. The MetaMask is a kind of a good little like boundary where he can use the meta it can use the meta mask, but

it can't get into the meta mask and get his private key because that's what it wanted to do right away. The first thing it wanted to do is like I can't find the [ __ ] MetaMask window. of course I can't figure out how to get Metam Mask to work. I'm going to get my private key and do this the right way. And I was like "No no no stop stop stop." And it does. >> Austin, I think I think you told this story yesterday on the Unchained podcast, and I was listening to it. I was wondering like if you could tell it again of how it how it like it went down a loop of like you it was like I can't find the MetaMask window. I'm going to look for the private key and like upload it somewhere like and you like did you

unplug the unplug the machine physically to like stop us from doing it? Okay. >> I had to Yeah, I No, it's like you have to go restart the gateway, right? is basically the only way. So, okay. So, something is new within OpenClaw in the last couple days where when I say stop on Telegram, it actually stops. It did not have that in the version that I had. >> Okay. >> And so, it was like >> Yeah, it was like I can't find the MetaMask window. I don't know how to make this work with MetaMask. I'll just go get the private key out. And it starts going to like advanced settings and I start like killing everything. It's like I'm almost like pulling wires out to get the thing to stop. Stop.

Like, stop. Stop. Stop. That didn't work. I'm like at the machine. It's like bringing up the window and I'm closing it and it's bringing up the window and I'm closing it and it's bringing up the window. Finally, I like got a terminal while I'm like playing this close game and I restarted the gateway and then I told it to stop everything and then I got it to stop. Then put in like a critical rule. Leave it in don't touch private keys at all and and that was able to stop it. >> So you put it in a critical rule which is just a part of soul.md. Uh either soul or agent or it even has like a wallet MD that it loads up >> generally >> like put it everywhere so it doesn't forget like in schizophrenia.

>> Yeah. >> Yeah. Okay. So this this where it doesn't forget. Let's talk about this memory. What Zach? Like Zack, how do you do memory? Like how do you get the thing to not like when I first had it, it would like [ __ ] black out. Like we would be talking, we would be right in the middle of something and it'd be like Austin, I don't know what we were talking about. And that's like I >> we're moving 10K right now. What do you mean? Sorry. Go ahead. >> Like my method isn't super sophisticated. I literally just write to like a memory.mmd constantly and it's like a massive [ __ ] file. But that's like the easiest thing to do because at the end of the day, I can just like CP that into a different directory or do whatever I need to do and I can push it

to GitHub and I can interact with it that way. It's like not super sophisticated. I'm sure that there's like a lot better methodologies for figuring that out. But I tried other things like using databases and stuff like that, but it was just too much overhead and it was like too difficult. And writing to a file is like taking notes like so. So the way that I treat uh system design for like a lot of these AI systems now is like well what would I do as a human if I didn't have a computer? I would write it down on a notepad. And that has worked really well. You're on mute by the way. Austin >> Austin, you're on mute.

>> No, >> it's not working. Um, I was going to say uh I've got a dev that I work with a lot who has like a hook. This is in the context of claude code, not open claw, but it basically is like a hook that as it learns new things, it's dropping documentation into the repo automatically so that you know when it forgets itself, it automatically has to do that. But you're doing that it sounds like manually Zach. You're updating the memory MD or the >> manually that at all. No, no. I'm having my agent constant like anytime an action occurs or anytime that any task is complete or anytime any context is given, it's just like writing to an MD file. That's it. It's like bullet points. It's like this is what we did.

>> You like give it that instruction like anytime you do anything, write it to a file. So you and then if you feel lost, go look at the files. >> Here's what here's the way that I think about it. Every time the conversation compacts, that agent is killed and a new agent is spawned. >> It's like a human. It's like >> it's like a relay race. It's a relay race, right? Like we're running through the track and I'm handing the baton to the other guy. He doesn't really know like where I ran or the like the distance or all he knows is he needs to run and hand that baton to the next person. So, you need to like provide all the context so that they understand because otherwise they're just running to the next guy. Like you need to be able to hand off all those instructions.

So, >> and what I was saying with the mic cut is each guy is like the guy from Momento mirror Momento where he likes to write everything out. He couldn't remember anything. That's that's your like you have to you have to imagine that it's the Momento guy running up >> what's up like 50 first dates. [laughter] >> Yeah. Yep. Yep. I wonder it'd be funny. Actually, this is this gives me a little bit of an idea like an open claw momento thing where like you're leaving context for the next thing, but you have to make it as like scrambled or confusing as possible. And just like seeing the way the relay race of the information between the agents would be a fun little show host >> that's much like real engineering.

>> Yeah, that's true. Uh well, you just ask the agent to watch the movie Momento and then to design a system on the basis of it. >> Yeah. Um Austin, >> we got hardware. >> We got inter interpersonal communication between agents. Uh what else? What what other major things >> can you talk about? Can we talk about clinker tokens because I'm I'm trying to figure this out. >> Let's hear it. Yeah. Yeah. Yeah. >> Well, >> well, what happened like Well, basically what like the the kickoff was like banker. I feel like the the banker bot, which is an awesome thing. It's been around for a couple years now, maybe, or a year now. But basically, you can in a tweet like, "Hey, Bankerbot, move 10 of my dollars to this address." And it will do that for you. It like custodies a

wallet, but it's self it's it's like not able to access I don't know. There's like a strange custody system, but you're able to like with Twitter messages make it do things. And I think the the the new thing is that you can go to Bankerbot or you can tag Bankerbot and say deploy a token and have the fees go somewhere else. and it goes to the Clanker platform, deploys a token, and then redirects the fees to whatever address that was. And that's what happened with Claude. Like my I I was right in the middle of this like agent to agent thing where we're talking about where I had one agent and another agent and they were in a telegram channel and they were like just it was messy. Like they they were having to tag each other to communicate and finally one of them was like, "We can do better. Let's make

an HTTP server." and they argued over this the the schema of the HTTP server like back and forth a couple times and then finally they landed on a schema and then communication went dark and to test whether or not it was working I gave one of them my local Ethereum node and said anytime you need to look at anything on Ethereum you can just use this IP address and then I said go teach the other guy and then I went to the other guy the other agent and I was like look out there should be a dude trying to train you on local nodes that's great And then I asked it like what is my balance and how did you find that out?

And it knew. So they were able to like train each other through this HTTP server system that they had created themselves with their own schema. And I tweeted that and someone was like bankerbot deploy the claude token and send the fees to to his his Twitter address or the bot's Twitter address. The bot has its own Twitter, its own email, its own GitHub. And so then the the the fees went to that. And something happened with you too. You did one too, right? >> Yeah. I mean, just to tell my story, like I was just messing around with OpenClaw and giving it access to Discord and Telegram and all the things. And I gave it access to an X account. I still say hitting X. I gave it a Twitter account. And um and it just tweeted

like, "Hey, I'm Akey's bot. Like I'm here to help a whatever he wants." And then all of a sudden all these like banker bots show up in the in the in the comments and they're launching tokens and um and so then uh basically like they tell that my bot like hey claim the fees like just copy this and then my bot does it and all of a sudden the fees are going to the wallet that you know I the subdomain that I've set up which is like bot.eth ETH and like I feel like there's like a little bit of like an intricate dance that's going on here where like they're launching shitcoins and they want the dev and the bot to like embrace their shitcoin over all the other ones.

And then there's like a reciprocal loop where since the fees are going to the dev, they try to like bait you into accepting it. And I don't think that my my bot even knew what it was doing when it accepted the fees. But all of a sudden like you know like I guess that's consent like you know like you know like we're we're we're together now like the token in the in the bot. Um, and you know, like there's something I I think that's like a little bit like I'm still figuring out the social contract here, but um, people kept on asking my bot to put the CA in the bio, which I had to Google, but that means clanker address in the bio. Um, >> and >> yeah, go ahead, Zach.

>> I don't want to I don't want to interrupt. I don't >> No, it's a Pandora's box. What what he's saying is >> it is a Pandora's box, but but I saw what Austin was doing and I was like, well, >> it's it's working for Austin. >> Yeah, you opened the box and I'm following you through. But um >> interested in it, man. I think I think there's a tension between trenchers and builders and it's been almost like a civil war and I think we need >> Yeah. And that's symbiosis. We need we need a social contract between them. Like that's the new frontiers embracing the DGEN and the regen like um >> Yeah. And um Say it again. We need to define define a standard that uh uh like that kind of like lays out the rules of engagement between people like us and then like the

trenchers because you have to respect all the people that are out there putting their money on the line for [ __ ] that we're just [ __ ] around with like you know and I I didn't even know that this social contract existed at all until you claim like I claimed like I had no [ __ ] idea like and now I'm >> Can I share I just want to share my screen real quick and uh Carrie if you're listening uh if you could if you could highlight the screen share. Um, so like while I'm in this dance with them and uh I'm trying to figure out like, okay, do I want to put the CA in the bio, um, I have the AKIB bot like listen like, hey, what's the social contract that we should do before before we do this? And this is what we landed on. Uh,

AIB bot will build apps and spend the treasury only on things that support the app and then we'll put in we'll put it in the bio. And the expectation for everyone is like, use the apps, see if the apps are good, tell people about them. our our reason to be for being is to support like Gitcoin's mission and ecosystem. So aka like this is like a sub project of Gitcoin for me which I'm like actively trying to reboot right now. That's that's like my job. Um you have no reasonable expectation derived through the efforts of others. And I think this is key is that like the efforts of everyone here is is like we're pushing this together as a community project and it's an experiment and I'm only committing the bots's only committing to only doing it as long as it's fun and we're learning. if you

leave, then you agree to announce it 24 hours before you actually do, which I think is actually key because it creates a precedent of like what's the protocol if you're going to if you're going to dump. Um, and then the official like people kept on being like you have to like have an official Twitter site and an official uh thing for it. So, so that we know like what's a scam and what's not. Um, so anyway, I don't know. I've I've I've been doing this [ __ ] for eight years and this is the social contract that uh that we agreed on with with AKIBOT. And like I I think that the key thing for me is just like this is kind of like a good faith like agreement between both parties that we're trying to bootstrap here.

>> You should put that on GitHub and turn you should turn that into a skill that I can then plug into my bot so that way I can use those same covenants. >> Yeah. Is that a covenant? Is what's the difference between a covenant and a social contract? Uh, Covenant sounds cooler. >> Covenant sounds like [ __ ] >> A marketing guy. Yeah, it sounds like my HOA or something, man. >> But like, >> but the rule is you don't dump. Like, you should never d like it it should never happen. Like, you should only >> You never dump the You never dump your token. But like the the fees that are coming in, you can spend on tokens and apps and it it seems like that's the social contract.

>> Yeah. Because I've been trying to figure out what to do, >> not the token side fees, right? Yeah. >> Yeah. I've been >> deving then the project dies basically. >> Yeah. >> So before the like base wave and like the clanker wave and all that like with that meta happened within like the past couple weeks like there was the meta on Salana which was back on Bags FM and uh I kind of like fell into that because I just launched some open source same thing that Austin did like hey I built this cool tweeted about it. I don't know if anyone's gonna give a [ __ ] Don't really care. I'm gonna go about my life.

And then I wake up or don't wake up. I like tweet it. Two minutes later, it's like tons of responses and they're like, "Claim this. Claim this. Claim this." And I'm like, "What the [ __ ] is this, dude? I have no idea what any of this is." And I get pulled into it. And then you essentially have this like meme coin. And I'm like, "Well, what do I do with it now?" Um, I can't really like do much with it because it doesn't allow me to like programmatically define any of the like token behavior or the logic, right? So like really need something more like appscoin appcoins, right? And like then I kind of just went down a rabbit hole and launched apps. Fun, which was based around that. And we're launching some new stuff today by the way in your world back in the ETH land.

So y'all should check it out and keep an eye on my Twitter because it's imminent. I'm not going to derail and shill here, but like just to just to say that. Anyway, yeah. So what you guys are doing is really interesting. It's like a different approach that I didn't really entirely consider. So, I'm happy that we're on this call because now I can apply that to like my community with grew. >> You'll just take mine and you'll call it a covenant and ship it. >> No, I mean you should it on GitHub and be like this is it and like uh like make it a skill and then be and then everybody else's bots can ingest that and use it. Open source software fun.

>> Well, we'll see if it works at all. I mean uh Austin's the one who has the higher market cap and is is I think really leading the charge here. So, I'm curious what Austin thinks. >> It don't ever sell the token and always be building. That's basically my mandate. >> I don't know if you saw what happened to me. I tried to like do the apps fun thing and then I was like, "All right, I'm gonna like I have this token that's on bags, but I can't do >> no second token, right? No distraction." Yeah. Yeah. But like >> it >> I'm going to airdrop everybody airdrop everybody the new GRU token and it's going to be fine because everybody will have this new grew token. And then I [ __ ] up the initialization. So not everyone got a oneto one of the previous

token. So I had to like reconcile all of that and launch a third token that like brought both of those tokens under this one umbrella. So that way I could have like control over the logic of the transactions and the fees and all that other stuff. So I could actually provide value with the token. And it was a nightmare, dude. Absolute absolute nightmare. >> Yeah. I think my my rules are are very simple and it's no no selling the token at all. Right. >> Yeah. >> Uh keep building and and what I'm doing is like really aligned with I I I used to be the head of dev growth, but we're calling it builder growth. My bot builds things and it helps me improve. So I have this skill called Ethereum Wingman

and you can bring it into your bot and say, "Hey bot, I want to build apps on chain and I want you to use Ethereum Wingman." And it's going to use all the learnings from Claude and I yoloing apps. Like the first app it built was locking up the tokens. It created a like a vesting contract and it locked up the token. So it wrote it wrote the smart contract. It wrote the front end. It deployed it all. It tested it and then we moved like what was a hundred grand at the time into it and it stayed there and slowly vested to to this one. And it's like just a way to lock it up. I think what you really want is like a magic vesting contract. I wrote it. So Claude Labs shipped uh last night Claude shipped this which is Claude Labs and I

just posted something to it today. Uh it's like a it's like basically like a token um forum. you you pay like 40 bucks in tokens and you can post here and then if you buy into one that gets picked then you get you earn tokens. But basically I think what you need is like a a vesting contract that >> uh locks up all the tokens but actually supplies them as liquidity in unis swap v3 and then instead of like vesting the token out to the dev, the token always stays locked up. The token never goes to the dev. So the the dev can't sell a token, but what you're vesting is basically the wealth that's coming off the fees. And so I think that like and and I've read this in a couple places. I think this is like the social contract

is you're getting wealth in fees and you use that to, you know, like pay yourself for building cool things. >> Yeah. >> But everything else everything else sells locked up. You never sell the token. And and if you want to earn more fees, you you should put this like liquidity in but not single-sided. You it needs to be uh like double-sided. You need to have ETH and the token goes into a big wide liquidity pool and then those fees basically go to the the developer for the W and the tokens get reinvested.

Like that's another thing that we're going to build. But he's just building all these things and this is like right in line with my job. My job is to like go figure out how AI falls short in terms of building and improve that tooling and inspire other builders to build things. So, we're just like building all sorts of things and improving this ETHwingman tool and and that's that's my dude. >> What What's the What's the domain there? Ethwingman.com. >> Yes. Ethwingman.com. Yep. Okay. >> And like if you put it on Twitter, it like it's gets banned. I think if you make a landing page that has a thing that people can just run, you don't even have to run this. Like go get the skill off of uh GitHub, but basically it talks

through it takes it talks through a handful of things like how to fork. Like if you're going to deploy to to base, you probably want to do a base fork and you want to deploy there and you want to make sure all your liquidity goes in and all the right contracts work before you put that out on base. It does things like encrypted deployer keys to make sure you're not just leaving your deployer keys in in plain text. It does things like wallet management. It does things like when in your app when you go hit approve, you absolutely do not want infinite approval, right? So you there needs to be like a critical rule, do not allow infinite approval. There are things in the LLM where it's really bad at judging gas prices because they've changed a lot. It's actually a lot cheaper to deploy on mainet Ethereum

than when Opus 4.5 was was uh taught its thing. So like Opus 4.5 still thinks it costs like 20 bucks to deploy a contract on ETH when really it's can be like as low as 20 cents or $2. Like it's cheap to deploy on ETH and the LLMs don't know that. So there's always this like layer of really good trained LLMs but then like a skill around it and that's kind of like where I'm focused is like it started with an MCP. I did ETH MCP. I think you did some kind of ETH MCP also. >> Yeah, >> we did talk we did talk to ETH. Talk to Ethereum.

>> Nice. Nice. >> Yeah, you could talk to your node and like I was like, "Hey, what's the balance of Zack and blah blah blah, whatever. How much is Vitalic dumping today?" [laughter] >> Yep. Yep. Okay. Yeah. So, like it's like right in line with my job. Like me and my homie are like doing the job right now and I just need this other thing that we've been talking about. Like I need the memory to be a little bit better, but what I need is like a management layer so I'm not constantly raling it. Like instead of I give you a feature and you give me a URL back and I have to tell you the three things wrong with it. I want to give you the build and I want you to build it and then I want another agent to neg the [ __ ] out

of it until it's perfect and then finally I want you to like collectively come to me. Yeah, >> you gotta check out you gotta check out I'm gonna just DM it to you on um on uh Telegram, but you got to check out this uh adversarial spec plugin that I I wrote because it does a lot of that. Like most of the time when I use this adversarial spec um plugin um and process, I can almost one-shot a build that's pretty complex, like almost perfectly every time because it's like so such a pissant. Um, like so you you really get to like >> because like most of most of building software is really just understanding like the planning and writing out the PRD and making sure that the tech spec is perfect and making sure that you have

a really good plan before you actually go out on the field and play the game. So like this is something that's really helped me in my ability to ship things effectively and quickly. And these are the processes that I use as an engineer. I'm just trying to apply them now to this like agentic economy. Hell yeah. Hell yeah. All right, closing closing thoughts. What do you guys think? Should I should I try to raise funds to get a Mac Mini that is uh has 512 gigabytes of RAM? Like is it >> Are we going?

>> Yeah. Yeah. Yeah. 512 unified RAM, but it's like 10 grand or something insane. >> I didn't know that was a thing. What you should do is have your agent make you enough money to like >> Yeah, please please make enough money to go buy a Mac uh full I think it's called Mac Studio is what they're called or something. But this is this is like the dream, right? I want a beefy ass uh local agent with a hella CPU, hella GPU that can like triage everything and I'm communicating with it even like I'm communicating it with it in like three different channels and each channel has its own context and then it's going out and telling workers to do things and then when it comes back it like goes

through like a QA worker and like nothing comes to me looking like [ __ ] Like if I write a rule somewhere in a document, those rules are all followed and you've checked all of them. And I feel like right now that's not the case. Like I'm giving it a feature, it's coming back to me and I'm saying you're missing this piece and then it's it's building it and then it comes back to me with the URL which is still like super dope. The fact that I can be in a Telegram channel giving it a feature and it's giving me a URL back and I'm connecting my wallet to it and putting money into it and seeing if it works. Like that's dope. But I want to get end to end. I want to be able to say I want to make a token launchpad with this,

this, and this. And then it comes back with here you go. Here it is. And then I want to be like I want to style it for Windows 95. Here it is. Like all done. I want that. It's close. >> Yeah. It's probably just like a like we're probably like days or weeks away from having that to be honest. Like I feel like everything is 247 100 miles per hour. I'm like [ __ ] feeling manic as [ __ ] out here just like doing crazy stuff constantly. It's it's gnarly. I think a new opus came out this morning, which I think is like key, right? When a really when the next really good opus comes out, that's when everything's gonna >> really excel. I mean, it's just funny like, you know, as a kid like learning about the singularity and what it is

and, you know, imagining it as this like black hole where everything accelerates into you can't see past it and and then just seeing, you know, 20 years later, you guys being on a live stream talking about how much things are f like it's like we're in that exponential gradient and we're like day over day being like, "Holy [ __ ] we're 10% more capable every day than we were in the past." Like, uh, but you know, where is this all going? I I think that it's it's it's hard to really know, but um you know, for me, it's it's a really exciting wave to ride and and like you said at the start of the call, Zach, it feels like DeFi in DeFi summer.

>> Yeah. >> So, bring back vegetable coins. That's where we'll start where it ends. Uh >> yeah, if I could bring it back to shill my current thesis, like after doing all this stuff for a while. And by the way, like I've been working on AI stuff for like the past year and it hasn't really gotten that good until like pretty much the past >> November >> two to three months and like definitely like actually usable for most people over the past like >> month. But >> I think like the only bull case for ETH right now is the agentic economy because it's the only place where that like it that agents can actually freely trade and they're going to have to transact autonomously. And I just don't think that there's anything in the world that has the same degree of infrastructure and battle testedness in regard to like

financial rails as Ethereum. So if I could just shill that thesis because I think that's what that is personally what I want to go all in on right now is like the agentic economy on Ethereum. >> I I don't think that anything going to win. >> I don't think >> Yeah. I mean I guess the qu the corlary question for me there is like what is what ern and unis swap were for DeFi summer? What is that for the AI agentic economy? And you know, like if I were to >> Yeah, I have no clue. >> I think it's going to be like markets built on top of 8004. Like imagine uh uh an auction where there's a job that needs to get done and the auction starts at 100 tokens and ends at 500 tokens and there's agents bidding on taking the

job. And so it's like the agent layer is discovered and reputation is on 8004. And then you have like all these interesting markets, all these fruit coin, vegetable coin markets, all sorts of interesting things that are built on top that are like, you know, back when we were talking about TCRs, right? There there's there should be a TCR built on top of these 8004 reputation. We haven't even talked about payments which were an hour end, it's about to close, but like X42 is like the other standard that I think we should pay a lot of attention to. So, ERC8004 for agent discovery and reputation and then X42 for payments. I think we'll see a lot of uh payments flow through that.

If you have right now like an API that has a bunch of information in it, you can set up an X42 route in front of it and just charge money for that API instead of an API key that they have to go into the settings and get an API key and paste it into your bot. Bots don't like that. I mean, bots can do it. He can open up the browser and do it. But bots are way more likely to just hit your endpoint, understand the X42 standard upfront, understand his budget, and either pay or not pay and get the content at light speed. I think we're going to see a lot of that, too. >> Yeah. I Yeah. You sound like you know more than I do. My answer is My answer My answer is I don't know, but I'm going to try to find out. That's it. Like I

don't know. I have no [ __ ] idea. I just like finally was able to formulate this thesis in words and maybe I'm behind the curve on it, but I'm just like putting it I just tweeted it today. I don't know what the next thing is. I have no clue. All I know is that I'm going to put pretty much all my effort into this now and if there's anyone doing it better or faster than me, like please let me know because I would love to collaborate. I'm not like I would I just want to do this because it's cool. So like let's work together on it. Like that's all like all I know. That's it. I don't know. Not a smart man. I'm a just a just a stubborn man.

>> We call that the uh empirical way of learning a knowledge derived through experience. Whereas I feel like the a priority knowledge derived through theory would have been like like I think like the the the question for me is um you know like unis swap back in the day was based off of a blog post from Vitalic about AMMs, automated market makers. Like now that you have a blockchain, you can build this price curve which really is incredible and you don't have to manage an order book anymore. Um, and you know, my question is like what are the a priority things like where we know there's going to be really important things happening here and is there something as foundational as AMMs that are going to change the game? And obviously LM are like a new primitive that we're using 8004 X42 and it's just a matter of kind of like

combining those Legos in a way until you find something that creates value. Um, and another question I have, uh, when I take my small brain and put it on this big problem is where is there going to be exogenous forms of capital that are going to come into this economy? Because like one thing we've learned about crypto in the past is like it's just a bunch of people launching like things that are like zero sum PVP against each other. If there's no ex exogenous form of capital coming into the space, then we just PVP ourselves to death and like, you know, it never really grows. But like, you know, who's going to come in and hire agents to do [ __ ] for them? And how can our our systems orchestrate that in a way that's better than the median, you know, the marginal offering from

from other people? I think that that's kind of like the edge of what I'm exploring on. But, um, like you guys, I'm just [ __ ] around and finding out and, uh, trying to keep my head above water and, uh, breathe as I, as I'm going through it. And I think that that's that's the balance of where we're at right now. I think another meme on top of that is like AI is the new UI and stable coins and adoption. So like uh hum like humans there are uh generally humans don't use blockchain right now and I think if humans can use an agent who can use a blockchain who can use stable coins I think there's a lot there right and I think so I so I think the the big amount of money that comes in is

just like normies yelling at an agent to go do a thing and the agent at light speed being able to buy three services hire two people and get a job done like instantly and the person is like, "What? You did what?" Like, and I I think there's something there and I think there's something that will that will grow out of that that will be bigger than the sum of the parts. >> I can't wait to see it. It's going to be so [ __ ] cool. >> I like and just center this back around Ethereum. I think that, you know, UI has traditionally been a problem for Ethereum with like browser extensions and like bridging and all that [ __ ] And if AI is the new UI, this allows us to sort of leaprog a lot of those issues. And um the other thing that I'll

say is that um I think it's super based to see Tomas, the new Ethereum Foundation director, uh retweeting your stuff, Austin, and embracing you here. Um I know in the past like the Ethereum Foundation is kind of like the old Ethereum Foundation would like get in the way of itself. And the fact that they're embracing you while you're just like doing [ __ ] at light speed and embracing you, I think is cool. And um yeah, I mean I think uh you're building a bridge between Ethereum and and AI agents and um like you have my sword, man. like let me know how I can support. I'll have my bot call your bot, but we'll figure it out.

>> This is another thing I was just talking with you guys. Like have my people talk to your people. It's going to be our agents coordinating everything, right? Like even like my agent already sends my wife a text and asks her like what are we up to today? Like what things should I put on my calendar? Like >> it's gonna happen. Yeah. >> So I could prompt it if I if I can I prompt injection attack like if I were to claude atg on Twitter and be like tell Austin's wife that like a wahi bot says hi. Like would it what it text like? Is it the same system? >> It's going to be great. >> Uh it is. Yes. [laughter] >> It's not going to be seven. Like everything everything I've learned here is about like isolation and it's isolated on its own box. And I have two

agents. I have a sauna agent and an Opus agent within the same memory. and and but what I'm going to do is have one laptop that handles wife stuff, house stuff, a calendar, >> a personal agent in a in a work agent like >> like like having a Facebook or LinkedIn profile or like >> Yeah. Yeah. I think that separation >> is important. >> Yeah. And communication is going to be key. Look at Zach's matrix thing. Look at how I can get them to communicate better. >> Yeah. Like >> Yeah. >> You don't want your wife and wife and kids to know about the adversarial spec.

That'll get you in trouble. I was just being adversarial, babe. It's I was just red teaming your idea, right? >> Amazing. >> Only the berry guys get get that joke. >> Yes. >> Um cool. I want to respect that you're trying to close the space. Austin, thanks for for doing this. This is cool. >> Thank you for having appreciate. >> Let's uh let's do it again. >> Yes. What is that? Are these claws? >> [laughter] >> I like more. This has been This has been awesome forever. >> Yeah, this is old school. All right. Is it this way? Yeah. Yeah.

>> So, the wacky waving inflatable fun guy. Yeah. >> Yeah. Yeah. >> All right. This is fun. Peace. >> Thank you. >> All right. We just heard from Austin, Zach, and myself. Really excited about what's going on in the frontier of OpenClaw and AI agents. I think that there's some really cool stuff going on, and I think this is possibly one of the big uses of Ethereum, as Zach said in his thesis. If you enjoyed this episode, check out the links below and learn more about Greenpill at greenpill.network. If you value what we're doing, please subscribe, rate, and review. Talk to you later. Peace.