Beyond Chatbots: Building Interoperable AI with the Agentic AI Foundation (AAIF)

Transcript

[music] [music] >> Okay, cool. Great to see you all today and here to talk about the evidence layer for agents. So, this is my first Devcon and uh I think you'll hear you've probably heard of some verifiable and tamper-evident records uh and what my goal is is to try to collaborate with all of you to have a standard that is portable and can be shared with lots of different agents.

Uh I think the evidence layer is important for the settlement rails, whether that be what's going on here or other networks. Uh just very little about me. I've done a lot of open source companies and projects. One was a company called Ohana that was based on open source Presto. If you've heard about the SQL distributed SQL query engine, that was part of the Presto foundation under Linux. And with this agent action capsule is the open source specification and implementation both at the spec is at IETF and these are four of the key ones.

And we're also a member of the Gentic AI foundation. In terms of the papers, really recommend checking these out and I've been checking out ERC 8002 and we'll be contributing as much as we can there. At a high level, we all know that the internet was initially started as links to anybody and you could be just like anyone else. But over time, it all the value coalesced to you know some companies that were the toll booths and I won't mention the companies, but you had things like search and you had things like

e-commerce and you things like uh social media, uh mobile applications. And so uh what was the the factor for them, the moat was having some type of uh trust associated with it. Uh if I make a purchase from a uh one of these retailers, I know that if I just give a credit card, they'll probably send me what whatever they're selling. But uh with the view of what's happening with agents, we're now at a a different a whole another opportunity to look at lots of agents per person or per company. And the question is, will we go back to yet again more of a centralized uh giant companies that have uh multinational agents all over the place,

or uh a world where we can have agents that are working with uh other individual agents, which might spread out and allow the uh communities to uh keep value a little bit more. And so that's more of a peer peer model. And the thing that's missing is trust of a stranger. So that's the whole talk. And if you look at uh a couple cases right now, the uh agents personal AI is uh all the rage right now. Uh you know, Open Claw was just really in November last year, and then Aramis, and now things like Muse, Thoughts, and everything else. And as

they work with other people or other companies, uh these are all uh interactions that can go onto a history. Uh and we look at evidence as facts. These are things that are sealed and uh able to be uh tamper evident. And uh if you look at the evidence and you can then put a a summary of them on a history card, uh that's where you can have the ability to make a designation about trust. Uh in our views, there's not any scores.

Uh this is something which uh every agent should be able to uh decide themselves. And this is not my vision. There's I believe that uh the Ethereum Foundation is also uh has similar views as well as the Project Nanda of the uh uh that's been around and uh we have a few other members of Project Nanda Project Nanda here too. In terms of uh these records, uh those uh can be uh uh depicted in a history. So every agent has a very small uh uh store which has all of the hashes and those can be verified uh for free uh offline.

And uh uh history card is a special type of one that has the semantics of what happened. And uh now the uh case of uh the readers, uh let's say for example, I'm trying to get not something from a e-tailer, but I want to get a DJ to show up at my birthday party on Saturday. I tell my agent to go find out. Well, that agent will go look around and of course, you don't know how much they're going to try, the effort. But uh they will be uh hopefully looking pretty broad, not just the uh the famous top three DJ uh uh service providers. But maybe there's some uh hobbyists and others that I

might like. But the question that they'll have is how do they know if I pay 10% of it down and they come to my birthday party that I'll pay them for the rest of it? Well, I could show them a history card. And that card would say here's the types of things I've done before, whether it be for hiring uh tutors for my kids or maybe it's for uh uh other ecosystems like selling um my Mac Mini for inference. Same way that DJ might say, hey, I am uh uh not sure I'm not sure if that DJ is any good. They can tell me they did all these gigs, okay? And I can look at that and ask for it. So it's very simple in terms of how that decision's made.

And as you have these intents that are made, I wanted to get a DJ, agree, pay, that's all sealed. And then the key part of this is getting the other side's half where they say here's what I saw, here's what I saw, and then that's where you can corroborate with a judge. And so this is how that works. And now you can share with protocol around evidence requests. And here's where we hope to collaborate with everyone. We want to donate this to a foundation such that you can now staple together history cards. And that's the evidence request I just discussed.

Today we have this in a couple places. One is a really cool open source project called Mesh LLM. That's out of the company called Block. And that allows you to sell your inference or trust another machine that you don't know. And the key part of that is all the things they say about their model and the the temperature or the quantization factor, those are all self-attested. So they use these libraries of the Agent Action Capsule to now get these to be witnessed. And then it can create a agentic economy for inference.

And then we also have personal agents. So if you have Muse or Hermes, you can just install these. It's a skill file plus a CLI. And the agent knows how to use a CLI to create these sealed records. So pretty easy to do. All you need to do is go look at agentactioncapsule.org. And I look forward to working with you. We welcome you to the to our community. We'll join everyone within here and all Apache 2.0 and Unencumbered. >> [music]