Agent Reputation, Delegation & the Return of DAOs

Transcript

Intro & inspiration from previous talks

Okay. Is this Kok Karok night? Oh no, wait. It's my talk. What's up everyone? Um, let me put this up. So, how many of you guys heard about the show talk from Eliza yesterday? Yeah. So he kind of like inspired me like he said why I'm actually doing slides and he kind of like sit down and started solo speaking. So I basically like started to I converted my my slides in a quick brief um and give context around a0047701.

Um and then we will have like a small like Farsa chat with James from Collaband speak about permissions and more. So quick quick check in. How many of you heard about 8004? Sounds good. 7710. Okay. The delegation toolkit like gators. Yeah. Couple of people. Okay. So, the TLDDR is I just give some some context here. Um yeah, I think like you know everybody's spun up uh his hope and claw. Um maybe you kind of like build up exactly what they were talking about around sub agents. Um but there are different like ways to do things right.

The agent PMF phase (early innovators)

I just want to run through like okay where where are we kind of like in the PMF and I think like we are obviously in the 2.5% indirectly what crypto twitter is showing you on the pipeline um we will we we are actually like quite innovator in that space even after like building uh multilow with n8 or testing out like trading boats. Um yeah so quite quite early I will run through some kind of like slides um and uh also like I want to give some kind of like context uh in terms of like you know uh this I like this meme and I'm using for everything but uh the TLDDR is that you know can profit from crypto trading can you know the question is you know garbage in garbage out um whatever

we are building it's important that we're training obviously with with uh with data and uh we are in different phase here um I'm just running through some slides because I want to really get you guys on the on the far side chat. Um so we are right now in the discovery phase where for a long time we're speaking about Jantic economies. Uh we see a lot of like testing being done. Um not a lot of like standards obviously like A004 is one of the main standard trying to kind of like attempt to build a full social reputation for those agents.

Discovery → Trust → Commerce

Um and I think like we are kind of progressing in the right way right we we see like we start seeing SAS completely like u um going uh yeah interacting directly with agents but obviously like the infrastructure today is not 100% on for agents right especially on wallets wallets is definitely like one strong use case because today humans are going through like um the interface your extension mobiles and uh there were like some experiment like metamas I don't know if you guys ever installed or you have some snap But snaps are actually going through uh without actually an extension is through endpoints through an API suite and it's quite interesting how this infrastructure has been built even uh even before um I just run

through some slides but basically like you know we we are like following like you know discovery trust coms and pay uh and now we are mostly in the discovery and trust phase um why trust phase because obviously we want to build a social reputation um and we want to profile Idos agents with specific like standards. How many of you guys heard that X42 for payments for example? X42 is is very strong for uh for agentic payments. Why is that? Because this is standard standardized um uh like a payment infrastructure. And that's kind of like the second phase, right? And not just operator need to be fully like verifiable, but you want to make sure they have the same instruments to um to kind of like connect with each

ERC-8004 explained (agent identity & reputation)

other. um and kind of like create specific uh transaction and uh one one things obviously like we discover pre- agents with for example Eliza there were at one point um there were like around 250 like Eliza forks uh they were basically like doing the same things those those agents going online right and I think on the consumer eye like things were very kind of like dangerous because you could not really attest a specific track record of this agent and I think Um that's why I'm I'm super I'm contributing quite strongly around a004. uh one of my colleague is is one of the the out Marco and and we are advocating a lot during this uh uh

this Denver this Denver time because I think where with uh obviously open CL is one one part of the metal but this entire narrative around agents right now is getting very very practical meaning also like we need to give the instruments but also we need to attest that those um that those agents uh are able to uh have a track record are able to have a social reputation but most important you can profile uh those uh those agents specifically on a on a on a common repo if not like we are not going into the second phase and the second phase is agent to agent commerce right um and I think it's quite uh quite important um yeah let me just uh go through a004 I think I mentioned a little bit what what

The problem: 300+ forked agents

is for there are three parts I definitely like recommend to see and read the um the ERC um but um but there are different the there are the discovery the trustability um and there is a third one around indexing but that's um that's kind of secondary here are some example um how is this done on the creator agency side and there are different things that will go wrong I think like in the first chat before like there were different things that mentioned um but obviously like uh um yeah without a proper like standardization around a commerce for example we're not able to um to trust the agent or even like the agent

discovery will be harder right um and there are different like use case gaming agent planning agents agents I mean right now we see around for LLM for example like the planning phase is super important and I think like having agents helping on the planning phase is is is very important but also like the discovery side for those agents is getting hard if you look at docs docs are being built for humans right uh right now for example we started to see docs with like Robbo text we started to see like direct JSON file that are um um direct inestable from uh from the agent side and for example monor repo build being built directly uh I think that's so things are moving forward but still like there is no for example common

Why infrastructure isn’t agent-ready yet

knowledge um for a specific use case right I think there is a workshop very interesting workshop about how the graph is building a um common graph based on use case for example I want to give I don't want to I want to give the agent a task and build a um build a specific like privacy uh DAP and right now like this this agent need to kind of go around through different privacy stack. So uh the entire kind of like GEO uh the SEO for agents is also like uh not made easy uh today. So the planning phase is is something uh super important and obviously there are like different uh use case today but they're also not new right like trading boats has been around for long there are executable on a

python script since the 80s uh and nothing changed uh much apart giving context I think uh um and yeah doctor agent is another example there are tons of things going wrong I definitely like emphasize about uh uh maintaining data on tees training model I really like the approach ical neon for example training data fully on private basis. So they're all different things. Um but uh but also we see like innovation going extremely fast and is not keeping up with like the problems right one example of a walkie kind of like sub agent happen uh I'm running like uh seven sub agents 24/7 uh and they're completely like on uh on

on the open right and what what is he said about prompt prompt injection is something like super super like important and I think today to tell you the truth like right now people are giving random permission to to basically like everything they're not really concernized um uh or sandboxed uh application because you can override souls um of of the agent you can override like MD file of restrictions and one uh solution out there uh that I've been advocating for a long time is is delegation is permissions uh and uh and it's funny enough like MetaMask has been working on something called Metamas delegation toolkit I don't know many of of many

x402 & standardized agent payments

people heard about this but basically like you can give delegation to do things and now that agents are around that's actually a perfect use case and I give you quickly like um a rundown what what is happening on the A004 to close this part uh the protocol is done in on also like on mainet since uh since last week and there are like 100 plus companies supporting a004 I think like there is an a004 repo with 20 plus uh thousand agents today so and uh And obviously like things are not like finishing now, right? There is an agent kit SDK where agents can actually support the same standard and uh and everybody here running agents should actually like kind of like align with this standard or have the agents align

with the same endpoints because if not you will not support X42. It will not support different kind of like endpoints. H okay. So second part of the the talk here is that how we get um to kind of like a gentic wallet. I just want to kind of like mention this part around um right now like wallets as an example is kind of like human uh first interface uh first and there are kind of like different different way especially on the friction side um that you can't uh you can allow a new UA to uh to run different things although today having like a separate UA a separate machine is probably the uh most uh uh secure way to do things because uh um I don't know if you test out other kind of like no

Why permissions matter more than private keys

interface wallets they're also like um yeah it's they're very nient right it's very it's very early stage so I mentioned what gator is right so you have both the delegator uh framework and the the toolkit uh is all based around 7710 so it's on the smart contract level so you can basically attest and give permission to do uh things and uh I will just run through a little bit faster but you have different kind of like uh parts you have the delegation You have the caveat enforcer that means you think about is some kind of like safe for permissions. You can actually like give permissions to the agent to do things in this in this example but also like to uh to to run different uh uh way. And then

you have the delegation manager that is kind of the um orchestrator on on the delegator side and there are tons of uh use case. I just want to run you through the uh some use case and not bore you with more um technical side. Uh give me two sec. So so hello Gator was was one of the uh example and you can basically like create a uh create a full gator app that is based on permissions and right now for example we are adopting this for for specific like agent infrastructure. uh the skills are already submitted in the main uh club repo that you can actually like build your own permissions. Um and uh some of the example for example like

Delegation toolkit (ERC-7710) overview

uh we did that uh back in ECC uh we run something called like EA red balloon where you can actually uh you will find like a red balloon but you can delegate the ownership to find this balloon directly to another like UA and the cool thing is that when the reward if you win uh when the reward is uh distributed automatically the um on on the smart contract level uh the uh contract address will be withdrawing directly to the the winners. So it was very interesting because this is not just a random uh talk but was all defined all the caveats were defined directly on the smart contract level.

That was one example that was very very cool and mostly like a test and different things like propel was also an example that we did on in defcon was a crowdfund uh cooking class where you can actually like crowdfund specific like classes on um on a different level and I think like at the end there are tons of like use case that you can actually empower agents for um for permission specific um allow UA to contract to have a contract code UAV gator super power. Um, think about also delegation of like a sponsorship. Think about agents paying for for gas for specific uh things.

Caveats & guardrails for agents

Think about uh account recovery and trust um and trust automation. Uh there are tons of use case. Uh one things I definitely like encourage is to check out like uh what uh what are really like permission and how you can enforce those on the on the agent side. if you are interesting to spin up your agent and not leaving your your private key prompt engineered the next day. Um I think that's uh that's definitely one uh um yeah those are tons of use case that I was having in my mind I dream about I wanted to talk about it but I think like I want to just jump in on uh on how James from Collaband leverage this smart smart account kit leverage this 77 um uh 10 like um permissions.

Yeah, advanced tracing is super interesting also like you can actually like define uh those permissions and what um the agent can do or cannot do and those are not something that are defined in in the soul or in the constraint in the MD file but also are um is something that uh is much you cannot basically like have it um um overwritten. Yeah. Because right now a prompt engineering like for example in open claw people are overwriting file or they're like backing up in their github repo and basically you can prompt engineer to take the the github and and check for example the private key and you know private key is just one one problem there are tons of problems right especially if you're doing like use case

Prompt injection & private key risks

like with uh like speculation specific right um yeah that's that was also like a very interesting u um uh permission sharing use case where you can actually stream token, you can build subscription streams. Uh but you cannot really like fake it because those were uh specific on um um permission level. So that's also some some example that I wanted to um uh to share. Um yeah, so that's it. So kind of like to recap like what we kind of learned today. Uh very nent phase. Uh I definitely recommend to everyone to test out. I mean open close is just one example but there is a bunch of web three like agents you can actually like test. Um but I think like

you know the the space is also moving around like permissions and delegation delegating the right to do something and there are tons of use case and uh you know we spoke about stable coin to be like one of the biggest use case in web 3 but I think for agent specific like permission is one of the PMF that we need to solve as soon as possible. uh if not will be kind of like a an agent fork war uh and uh is not good for like consumers right where consumer actually wanna I want to consume agent and do specific things okay I think we're good um I think like let's let's invite James here on on stage so basically like James with collab leveraged let's so he basically like leveraged like 7710

Smart accounts vs EOAs

and the entire permission uh Um maybe you want to give like some colors on what is Collaband and why this is needed for for agents specifically? >> Yeah. Hi everyone. My name is James Young. I run a project called Collaband. We do what's called token gating on Discord and Telegram. We've integrated uh the MetaMask delegation toolkit uh 7710715. Um and we don't have a lot of time so I'm just going to do a quick check. How many people here are developers? Okay, so hot take real quick. I'm going to speedrun some of this. You don't have to have your agent manage a private key.

Those days are over. You have an agent with a smart account. You use uh 4437 uh to have your agent own a smart account. you use the MetaMask delegation toolkit to create delegations because now we're in this world where I think personal agents or having an agent do something and having that one agent then orchestrate to sub agents or what is known as a swarm is going to happen. And uh there are so this this primitive where you have these agents

Agent swarms & high-frequency transactions

24/7 doing high volume transactions. You can't have a swarm of agents first sending funds to each other in order to do work and there's just so many transactions. And so what you do is you have a smart account and this agent has a smart account. You issue delegations. And what a delegation is simply is a IOU. It's a signed message. It's offchain. You store it. It could be a ZK proof. It could be in the open. Whatever you want. And you then have these agents have all of these IUs. But then how do you know if the agent has successfully

performed a task or how do you provide guard rails to an agent? There's no way to provide a guardrail to an agent if you give them an EOA. So this is where the MetaMask delegation toolkit with 4337 comes into play. You have this notion of what's called caveats and Franchesco talked about caveats and caveat enforcers. These are guardrails so that you can actually have the agent perform the task, give this IOU, and you can enforce to see and check if the agent did what it said. Now, this doing what the agent said it did can be logged into a reputation contract. And this is where 8004 comes into play. So, we have

Off-chain delegations (IOUs) & enforcement

all the ingredients. I know that Griff and Aake uh were talking about the DAO and liquid democracy uh and you have this chain of delegations and whenever the DAO hack happened I don't know what is it now almost nine nine 10 years ago almost um these primitives didn't exist we have AI agents now that can perform transactions at scale and all of the ideas and nent ideas that we experimented with and I really appreciate this kind of talking about failures uh from the previous fireside chat because we can take those learnings and now we have the primitives to

actually do it again with the learnings from before. So I imagine and this is the hottest of hot takes. I think Dows are coming back and they're going to come back in a huge way but not how you think. I think agents are going to help organizations be able to be more efficient and really uh agents I believe if you have this hybrid organization will be able to keep what I'm calling the DAO culture intact. And so there's a whole new level of experiments. Ethereum was able to solve this hard trust problem. you had uh all of these um

Logging performance into reputation systems

metrics, TVL, token lock transactions, and we thought that these metrics were going to give us what the health of a DAO was like, but we know that that's not true anymore. We know it's all about reputation. It's all about people coming together and how much you can trust the organization and the people around it. Yeah, I want I want to say James like we did bunch of like uh live stream together and I think like uh with collab for me was the first example that you actually cared about the entire social coordination and uh when we spoke about social coordination how this is evolving is this economy of trust but also like then we kind of analyze where are the instruments right and this was just before like the entire kind of like

agent narrative like started right even a004 was kind of like the main idea it took like one plus year to get uh into market now. It's it's it's kind of it's not a combination but I mean being on on mainet since like in less than two weeks it's also like crazy and it's coming like in a perfect timing and how we build this economy of trust is obviously like the do you have like reputation and you have standardization I think x42 for payment is is helping a lot but also it's kind of like a chicken problem because we have 20,000 agents supporting a004 and you know the majority of agent developer today they need to know that this standard exists. in order to operate. So it's also like maybe a

DAO lessons from the past

question for you because you've been around for for a long time around this social coordination is that you know how we push uh basically agent developers or even agent itself for the discovery to uh to leverage this um this standard uh and you know what are kind of like what are the scenarios here because we could have a scenario that uh is finishing up like an Eliza um 300 forks of agent doing the same things without not knowing what are the performances or we kind of finish up the scenarios that the top services will automatically support X42 and in order for the agent to uh to create commerce to have commerce you need to support

those similar to how mold books now I mean don't don't quote me I mean in six days the API are down on mold books but mold books force the agent to uh to um uh to basically like support some of the standard in order to um uh to intervene in and uh and uh and do post or engage post in mold books right so I think that's kind of like the two scenario that we see >> yeah so uh to boil it down um with Dows and communities uh when they come together it's all about trust now you also in the AI world we're seeing okay AGI is going to come and all the the doomers are saying uh you know Skynet and all why is because

Hard trust vs behavioral trust

we don't know if we can trust trust this AI. So you have a group of people that want to come together and figure out how to trust each other. You have the AI industry trying to figure out with AI safety and all of that. How can you have uh an AI you can trust? So I think the it boils down fundamentally to a trust problem and trust is so hard online body cue body language all of this things that you have IRL uh and you know trust is how businesses operate trusting each other. Uh so I think that we have this

trustless substrate to now reconfigure trust on and I think that is going to be this intersection between crypto communities and AI. Uh I think that this is a this is this is going to be the pivotal point if we can figure out signals whether it's reputation identity the these are big rabbit hole concepts uh that we need a lot more experimentation on uh I mean in this last year um you've seen MCP AP2 8004 X42 you've seen all these like standards emerge so we have the ingredients And that's that all what I just said

Trust as the core coordination primitive

happened last year. And now we're in 2026. You have Open Claw, you have all these things. And it's it's not slowing down. I think this is the slowest it's ever going to be. >> It's just going to accelerate. Um, but I think we we have these issues, these harder soft trust issues. Now that Ethereum has kind of solved what I'm calling infrastructure hard trust. there's these behavioral changes that you can't just look at onchain. And I think because we have all these new onchain metrics, it's very easy to look at them and say, how do we use these metrics, these market metrics, not not

community building metrics, but market metrics? How do you map that on to uh community building? Maybe you don't. And so this is where I think AI comes into play. Um I have a lot of ideas. We're doing a lot at Collaband uh to experiment with this in a low stakes environment. We have 60,000 tokenized communities. We have millions of people using Collaband. Um and I think that the new form of tokenized communities is the new we we have a glimpse into what does it mean to have these new online organizations

uh maybe it's not VC funded you know because it's so easy for anyone a group of people collectively to create apps now like with vibe coding code is pretty much disposable nowadays it's easier to just redo something if you've modulized and componentize the architecture to throw it away and just start again instead of having to refactor. So there's new paradigms and I think that we should all take a step back take a look at these new primitives that we have and try to go back to first principles. Yeah, I like what show and we have two minutes but I like what show said yesterday that you know like when is open source now you can you can literally copycat like different features that you have in my product but I think what it comes down is on the distribution side right I think like

Agent digital twins

people will build a copycat product and whoever distribute the better but how distribution is connected with the end consumer the end consumer need to have this kind of like reputation if not cannot do a specific like decision- making uh and this will same will be the same with like agents agent and discovery will also happen by by reputation. And I think we have only these instruments. I think there are like still some specific things that we need to kind of like uh uh get get out. Uh uh but uh but we have everything to kind of like build a strong economy. But one things to say is that uh permissions is something that we need to solve as soon as possible because if not like right now is already the worldwide west

with open claw. think about when actually like you start to have like there are already some but open frameworks for for web 3 specific right you have like Austin speaking about like wingman I don't know if you you test it out is very very cool um but yeah >> yeah I think EOAs are kind of like in the I don't know 50s and 60s of compute you just had root access uh in your laptop now you you don't you you have permissions and the same thing is going to need to happen when it comes to agents. And so now we have all the ingredients. And I think that the kind of idea of build it and they will come is no longer true because everyone is

Coherence scoring & prediction markets

building right now, right? It's so easy to build. So I think distribution also like you had mentioned is going to be huge um because people are going to fight to keep their attention uh because it's just going to be overwhelming. I mean you see all the AI slop right now and it's just going to get worse. >> And we see also like more players kind of like consolid consolidation, right? Especially in DeFi, right? Like bigger player get bigger. Uh yeah, so it will be harder to kind of like compete. I I saw like a question there. The other question maybe shut it out like we have like >> uh really stoked about Dows coming back for agents. Uh would love to collaborate. We're thinking about doing an agentcentric governance model for Boulder and spinning up server. Want to

bring people into that. I'm curious what you think are the missing primitives or design constraints to be able to actually create agentics forms. Um whether that's like existing primitives but configured in a particular way or where do you see like the the need for innovation or or creating new things that make those possible? >> I'm going to give a quick hot take on uh that because I do have an answer. Um, and I I I kind of alluded to this yesterday in my talk. Uh, you have an agent, your digital twin that represents you. The agent plays a prediction market game with you. So, what it does is it creates these scenarios and it gives an

Agents aligning with each other

answer to what you would choose in this multiplechoice quiz based off on the scenario. And if the agent corrects guesses correctly, your coherent score with with the agent increases. If you if the agent guesses incorrectly, the coherence score goes down. And then there's this threshold where what is the percentage of correct answers that it gives so that you can then begin to trust your agents. I think that there'll be Dows that are just agents, but they're steered by humans.

So instead of a Dow looking at TVL, looking at all these different things, you look to see how coherent your agent is with you. And this is AI safety. It's the same thing. It's all boils down to trust. Once you have coherence with your agent, the agents then try to have coherence with each other. Once the agents have coherence with each other, they are managing the DAO together. And you have now Dows that have coherence with each other. you have this network effect, but really it boils down to alignment with you and your agent. I think there's a lot of experimentations there and I think from what I've tried to research and study, we have all the

DAOs returning in an agent-powered form

primitives for that to happen. So now it's just experimenting uh w with with different scenarios that uh like that's just one example but I think that there are many more to come. >> Cool. I think time up. Uh thanks so much everyone. Thanks so much for bowler for for inviting us and yeah shout out for you guys to stay and check out this talk. >> Thank you. Thanks for