Demo: Cambrian ERC-8004 Data Agent | Ricky Esclapon at Agentic Zero

Transcript

Okay, great. Hey everyone. Uh, I am Ricky Esclapon. I'm a senior data agent architect at Cambrian. Um, I work for Sam Green. Uh, if you were here for his talk earlier, you already know that we're uh super at the super interested always being at the forefront of Agentic Finance. Um, and we've been hacking quite a bit on the ERC8004 uh standard. Uh, just kind of experimentations and learnings. And I'm just gonna kind of quickly run through some of them and then kind of give a bit of a demo of some interesting things that we've kind of been working towards on this.

Um, so yeah, starting off, uh, we already had our an agent that we call deep 42. Uh, our agent is really focused on kind of bridging the gap between onchain and off-chain data. Uh, so you can think like connecting um, smart contracts to GitHub repositories and like social data. Um and also just generally maximizing the data that we have available. Um [clears throat] uh but practically speaking uh you can use our agent here. Uh we have our docs. Uh all the links will be on the at the end on the slide. Um but it's yeah pretty standard agent. We have an API product. The agent is just effective at calling on these different API products and you know giving reasonably good

answers. Um so of course you know uh the first step is of course deploying it as an ERC 8004 agent. Uh not going to spend a ton of time on this but I will say uh if you are looking to build these agents for sure spend a fair amount of time on the A2A standards uh the A2A agent card. Um it's pretty much all you really need uh you need to kind of get started. Um, and once you kind of figure that out, um, you can, uh, use the agent zero SDK created by, uh, Marco Marco De Rosi. Um, which is just fantastic and we've been contributing, uh, to it as well. Um, but the A2A agent card, uh, a little bit more challenging, but yeah, I'm not

going to keep talking about that. Um, the next thing we kind of looked at was kind of, uh, you know, what can we do with like TEES, like what what can we kind of improve on? This wasn't an area where like we necessarily thought we're going to get like a ton of ROI right away, but it is an area that we kind of wanted to familiar familiarize ourselves with a little bit better. Um, and again like so our end state here was you know we we made some uh steps towards kind of uh removing some trust assumptions. Um so like the code that is claimed to be uh run actually running. Uh there are still some trust uh trust assumptions though. So even like with uh this kind of TE model that we set up um basically

we had uh one TE for agent and one for an MCP server uh and essentially these are kind of like both talking to each other and then we kind of built this UI where you could kind of see like the proofs um and kind of like the whole process kind of running through um the yeah the only thing I want to mention here is uh if you are hacking on like TE stuff we just open source this repo uh so you can feel more than welcome to like kind of uh take that as like a starting point. Uh the other thing I'll mention is if you don't want to kind of go to the level of detail uh of like doing all this stuff yourself in GCP, uh for sure really recommend I en cloud uh compute, sorry. Uh super good uh for like deploying tees. Like I don't think

I've really come across like a better CLI tool. So if you're just getting started, I recommend that. Um it is a bit more expensive in my opinion than like maybe it should but uh for sure a very very good product if you're like looking to start looking at TE stuff. Um and yeah so more recently we've kind of been looking at the agent discovery problem. So again, like we we were looking at all sorts of things before that I'm not going to touch on. Um, but right now there's kind of a strong um um a strong movement kind of towards figuring the uh agent discovery problem out. Um because you know we all want millions billions of of agents but you know how do you like actually kind of like get to the data you need um do it

cost effectively efficiently like how do you kind of maximize on that? Uh so we were experimenting with kind of a new API uh endpoint within kind of what our agent does uh which essentially allows you to kind of plan like your optimal execution of uh what agents you should call on to kind of uh accomplish a given task. Um so at a high level um like I think knowledge graphs are like really an amazing primitive. Um so when we kind of look at it this way um essentially you know we're going going from like an original question uh down to capabilities so like A2A skills of the agents um then we so from the question basically we want to find does that

match in capability right like to the question and then from there uh you know we uh we have like a certain number of agents uh that we sip through uh we look at feedback and then kind of you know you can um turn that into like a routing plan and then like do also like the execution side of like uh kind of trying to optimize on these things. Um so I was going to give a demo. I didn't have time to or I didn't plug in my laptop. So hopefully this works. Um but essentially here is a demo where I'm asking you know uh which agents can uh can tell me about movies. Um and it's essentially going to kind of pull up the like relevant knowledge graph of like the information.

So I'm going to start from my query which is you know which agents can tell me about movies and then there's a skill uh associated to movies and then that's connected to the agent and then from there because it's in this like knowledge graph format you can do like pretty interesting stuff with like uh grounding your answers and like get like a so you see here there's like sources that are like associated with like the confidence level that the agent had and then from there the next thing we did is kind of turn that into like an execution plan. So as I said like the API we're like experimenting with uh kind of creates this plan and then you know this is just a more simple execution of of one of one agent. Um but you can do a lot and actually uh that's the that's

the part that I we still need to work on more is kind of uh making it more compatible across the board. Like there's a lot of u like we have standards but in practice like it's not that easy to kind of have an approach that works across the board. Um um yeah and a quick word on like uh knowledge graph databases. I think um so this piece of code right here so we're using a back end called Google Spanner which I very much recommend looking at. I think it's like an amazing database um because it combines cipher query language together with SQL and it's like in the same database um and I think it just works so well and in the past I was kind of um wary of like um like shifting to like a a different kind of database

from like past experiences I've had with like a gentic DBs. Um but this one I very much recommend and and here it's kind of interesting to just look um how simple it is like you you find your like relevant knowledge graph and there's many ways of doing that. Um and with knowledge graphs you can even like represent relationships on top of it like just content embeddings you can do like really interesting things. Um and here we're pretty much just feeding it like the the graph uh of information and then like our question combined with LM. Um, and yeah, we can get to like these really like good rounded answers just because like our data kind of lives in this format. Um, and then from there we thought, you know, how do you actually do this same thing but like in a permissionless way, right? Like I can't

I can't go and like, you know, sell like some version of Google Spanner u to other people. In fact, it's quite the opposite. Like u you're usually locked into like really huge things that you don't really need. Um so we kind of thought you know like what would it look like to provide a database that has like the same kind of capabilities uh of of spannograph um but you can actually use like more effectively in like a agentic environment. Um so this is kind of here yeah so this is kind of like a a very brief demo. So here on the right was like my first kind of interface and experience. Um, so here we don't have like the uh 402 payments yet, but basically what I'm doing is I'm

uh I'm just like creating a new account on the database and then I have like my own private view and then from here I can run like SQL or cipher kind of interchangeably. Um, [clears throat] sorry u. So yeah, here I'm just going to like basically start writing some data. So I did a query that had no no results because it was a fresh database, but now I'm going to write some data. So um yeah, basically data is in there. I'm going to go back to like my interface and now like okay data is there. Um so basically kind of work working towards this idea of uh sorry here let me um yeah sorry I'll skip ahead for the interest of time

here. Um so yeah basically like uh before I showed a demo that was kind of like uh you know we ingested some data from like subg graphs and like onchain information about the agents and then we have like this QA process uh that we can do on it that has like very grounded answers and then from there we can do like orchestration on it because we already know like like I'm very confident this is like the best agent to call on and all these things will like for sure get more complex like with things like reinforcement learning. Um but yeah, so for now it's basically like you know can we recreate the same thing but you know not uh kind of doing it from from our own DB that like we're building. Um and then yeah here's kind of an example of like just importing the the subgraph

data um like into the into the database and yeah it just runs through uh so you can kind of see like you know number one provision the database and you ingest it then you have like this Q&A process and then you arrive like orchestration uh kind of end result. Um, and this is like the one area I thought I think is like super interesting is like when you're using something like Google Spannograph, it doesn't inherently have kind of any capabilities around like um attestations or anything like that. But with our database like um basically what we can do is we can not only make it permissionless and you know anyone can just like go on chain pay for whatever we're using like make an informed decision of you know should I keep using it stop using it but you also kind of

have all these other benefits like uh so here's kind of a really ugly long list but um yeah we're just like a lot a lot more you can do when you kind of uh control your stack uh and you can kind of see that also here as well like I can take all the at the station info and kind of like do verification on top of it and kind of see uh exactly what's going on. Um but yeah, so like yeah, I I'll share some links here at the end. Um but pretty much like anyone who wants to like just try it out and hack like it's pretty much free and um and yeah, like and we have like a an agent that like essentially you can use to like provision a your database. Um but yeah, thank you