Unlocking Agent-to-Agent Transactions From Andrea Muttoni - Story

[Music] all right hey everyone both in real life and online my name is Andrea and I am from story and uh appreciate you being here I know things can get hectic and there's a lot of talks so today I want to talk about agent to agent transactions uh not really going to focus on on any specific ecosystem this is going to be a very generic talk um if you're into AI agents this could be interesting for you and um and yeah let's just dive in so before we start I want to set the context of uh why we are thinking a lot about AI agents at least at story uh because if you think about it agents are intellectual property both the inputs that go into training an agent both the outputs that come out of an agent and the model itself that or the M or the multiple models that power an agent are intellectual property and and so you know everything that they create is intellectual property everything that they know how to do could be intellectual property and I'm just setting the context here slowly agents are getting smarter and smarter if you think about how AI agents are today especially in web 3 uh I'm sure you probably will agree that a lot of them just end up being a reply Glide on Twitter right the most famous example is a BT and I think it's a great example I think they they're doing a great job um but a lot of agents are just thin rappers around open AI anthropic grock and other Foundation models and while it was very impressive and it created a very you know strong speculative uh bubble a couple months ago the novelty of just being a reply guy is is really wearing off quickly uh and while this micro bubble is kind of popping the macro trend of Agents is just getting started um and if we look at the current trends even in web 2 it'll help us think about how where agents are moving towards because every major AI company is moving towards agent first experiences we see this with open AI moving towards operator and GPT 5 is going to be basically a wrapper around multiple models and it's going to arbitrate and and orchestrate different models based on the task um things like deep research right Hardware companies like Nvidia are creating essentially supercomputers that you can buy for a couple thousand dollars and keep on your desk so creating specialized models is going to get easier and easier and cheaper and cheaper and at the same time you have these closed Source monoliths you know that have massive resources but at the same time you also have open open source models which are catching up quickly whether it's deep seek stability AI um and all of these three macro Trends are really pushing in a in a direction where an agent is not just going to be a prompt around a model if you create an agent on Eliza today really what you're creating is a character. Json which is a fancy file for a prompt um but we're moving towards proprietary model proprietary data think of you know maybe medical um agents that have access to specific medical data that no one else has access to um and and given that they're going to become proprietary they will have knowledge and knowhow that can't be replicated by just changing your prompt by changing your character. Json um and what does this you know create what does this cause is that agents will be increasingly differentiated and increasingly specialized and why am I you know hammering home this concept it's very simple uh when when agents become increasingly differentiated and specialized and not being able to easily be copied different agents will need to be will will need to lean on each other for different tasks so keep that in mind as we move towards the rest of the presentation um my personal opin opinion of of the phases of AI agents going forwards is in you know last year and this two years ago and last year we were kind of Awakening agents they were again prompts around simple models this year next is really going to be about learning and and becoming differentiated so being able to really decompose tasks and perform those tasks with specialized agents and we're going to see the start of a skill economy I don't know how many of you watched Matrix I mean it seems impossible to me that people may not have watched Matrix but you know I'm I'm also noticing that I'm starting to get white hair so um if you have watched Matrix you know how you know Neo or or Trinity was able to implant specific skills when when they needed them this is really going to be what's going to happen with AI agents as well um but in order to get those skills there's going to need to be a transaction and again this is the whole purpose of of this talk and as these transactions emerge and will become more and more commonplace a whole economy around agentic skills is going to form so um that will also bring with it reputation systems that will also bring with it um you know more fine-tuned performance right now agents are kind of like open-ended stochastic you know non-deterministic things that you kind of just let loose and see and see what happens but in the future the the uh models will also evolve to be a lot more deterministic and therefore a lot more dependable and this is the phase that we're calling ascending so as I said agents will transact to exchange skills and skills is really the key word here because transacting is easy and we'll see why um transacting on something that's a bit more complex is going to be is going to be difficult because communication is difficult communication is probably one of the most difficult things that can be done uh with humans let alone agents especially natural language communication um and if you think about how we communicate with natural language right I can speak to you and hopefully you're understanding more or less what I mean but what your understanding is probably a little different than what I'm saying H but it's very flexible right I can choose my words you can choose how in how to interpret them and more or less we're going to find each other in what we're trying to say um and for agents it's very much the same everyone can use a foundation model to generate natural language and everyone can use a foundation model to parse natural language but it's very expensive in tokens to continuously exchange information via natural language because it's very inefficient and as I said said earlier it's non-deterministic you don't really know what you're going to get both in terms of output and in terms of understanding the old way of communicating for programs was apis and the the great thing about apis is that they're very specific they're very deterministic uh but they're very inflexible they're very rigid you need to implement the exact API call in order to get the output so what if there was a way to merge these two things what what what what if there was a way to have the benefits of natural language but then you know kind of transition into an API form of communication and really if you think about it right now there are there is no Universal standard for agent to agent communication but people are working on it so Oxford University in a um in a research lab with Microsoft is working on something called Agora Agora protocol it's a web 2 protocol and um what they're they Hy hypothesis is Agents should start communicating with natural language but should transition to basically generated apis that both agents agree on so you start with natural language you negotiate a common interface and then both sides both agents Implement that interface and continue transacting through the API that they created themselves for themselves and between each other right and this is a very interesting concept and they they they have the whole research paper I think they're still in the implementation phase and we're actually working with them but just to give you a clear you know demo of it let's say a weather agent uh or a news agent needs to know the weather from a weather agent and they're two different models they don't communicate and they're just interacting with natural language um they they'll start off with natural language but then they'll they'll agree on okay I know that I need weather data you know that I need weather data so let's just agree that I'll send you this type of request and you give me this form of response right we both implement it and from from here on out we're just going to use the API directly so you save tokens you save time you turn it into a deterministic type of output and it's a lot more dependable but you have the benefits of starting off with natural language so communicating is hard but transacting is actually very easy if you think about money it's one of the easiest forms of transactions possible money is very easy to transact with and it was you know the first ever use case used for AI agents because first of all every blockchain was built for transacting on fungible tokens right um and the first web 3 agents the first thing they did was create an EA wallet and start to transact tokens back back and forth uh because the transactions are are Atomic and once you receive the funds given that they're fungible you can do whatever you want with them you don't need to have any form of meta agreement around how you can use those funds um so that's not the same for intellectual property and as I said earlier everything that agents ingest and everything that agents output or can output is intellectual property and the difference between intellectual property and regular Financial transactions or regular asset transactions is with intellectual property comes terms comes conditions I will give you this IP or you can use this IP but only if and you can set the terms right you can say you can use it but only if you know you use it in this context or only if you pay me 10% of Revenue of anything you generate and so creating those terms is is quite complex especially if you don't have a way to do it programmatically um because I first need to get permission to use it then we need to agree on a certain set of terms then I need to pay back anything that I owe and I need to keep tabs on what's on what I need to you know keep track of and then any derivatives created off of the IP that I licensed also need to abide by all of those terms so it's a compounding problem and in the real world I p is also complicated and uses a lot of meetings a lot of lawyers a lot of back and forths a lot of Google docs or a lot of Word documents but what if there was a better way and I said at the beginning of the presentation I wasn't going to you know shill any project but you know I kind of have to so um the good news is story that's all we're focusing on which is how to make IP programmable so all of the terms that an IP has are programmable onchain parameters embedded in smart contracts and therefore agents can transact on them can create condition Logic on those and compose on them but that's not enough and this is why I recently wrote a uh white paper around a transaction control protocol specifically for intellectual property uh which I published an AR on archive if you want to take a read but I'll give you the very quick tldr basically it's a trans agent to agent transaction system specifically for intellectual property and the the tldr is this diagram which is whenever an agent requires some form of information or IP from another agent that's not a simple Atomic transaction but has certain conditions attached the provider agent can generate and negotiate with the requester agent agent specific terms right and the way that they do that is via a common license format that I'll get into and they have this negotiation process back and forth then when they agree on it that specific set of information that is requested is registered as intellectual property and those specific terms are associated to that intellectual property and the license token is minted by the requester and therefore there's an immutable proof of what was agreed on and what was transacted on and what makes it possible is the programmable IP license so it standardizes commercial terms and metadata um we basically worked with several legal firms to create a standard license which we're calling the programmable IP license and all of those terms are parameterized um on Smart contracts so the agents can programmatically define those terms without having to parse legal documents and agent tcpip is already being used in the wild um so Luna is already using agent tcpip Through The DaVinci agent that is acting as her agent uh within story we had a couple experiments around um using agent tcpip to license training data across different agents to create a new agent basically merging training data of officially licensed training data and generate a new agent from that generating a new agent from that we also had a hackathon a couple days ago specifically focused on agent tcpip I don't know if any of you were at the super agent hackathon if you were uh congrats um where basically agent TCP was being created and Incorporated across all the major agent Frameworks and recently I don't know how many of you are in the AI space or know about the model context protocol but anthropic which is you know a very big web to AI company um has created the standard called the model context protocol and what this does is standardizes how specific basically units of knowledge can be plugged into by different agents right so the hosts which are the applications can hook into local or remote servers and clients that have specific knowledge basically going back to the skills that we were talking about um and the great thing about this protocol is that it's completely standardized so anyone can contribute their specific host their specific basically plug-in for agents to use and and we implemented one specifically for agent tcpip so the way that it works is the AI agent interacts with a story Model context protocol that exposes all of the IP related information and all the IP related actions and the agent itself doesn't need to know anything about IP so you can generate an image you can negotiate the license terms uh between the agents you can mint it basically go through the whole agent tcpip flow that I showed in the state diagram but actual live um I don't know okay so this is a quick example see if um let me make it go faster and press play oops okay so right now an agent is requesting a specific type of image from another agent right and the other agent is starting the creation process and it generates the image and let's say that image is a form of IP right now once it's created the agents are negotiating on the specific license terms that that that image now has right and they're negotiating on them they're agreeing on them and then it's being registered on story and this is the actual story Explorer where that IP has been generated that specific image from this demo um with all of the metadata and all of the license terms that were negotiated between the two agents so what's next at least for unlocking agent-to-agent transactions with intellectual property there's two big areas of further research that would love help from everyone in the community the first one is negotiation robustness this is also something that with the Oxford team we've been talking about that they're really interested in agents you know we kind of assume that all agents will be benign and honest and um but the reality is there might be just as in the real world there's scammers and and badly intentioned people there's going to be very bad agents going around and in an agent to agent economy those types of Bad actors are going to be especially problematic especially predatory agents that have very good negotiation skills for example right so working on negotiation robustness between agents is going to be very important part and making sure that the negotiation process is fair and the second one is moving ambiguity because going back to what we were talking about anything with natural language implies ambiguity I can tell you something I've been talking to you for 20 minutes and you're Pro every one of you is going to walk away with a slightly different version of what I said and probably none of those versions match with what I have in my head right and a little bit a little bit depends on how good I am at speaking but a little bit depends on just the natural ambiguity that comes with natural language and so how can agents start to really remove that ambiguity and we actually see this organically I don't know if you saw a pretty viral tweet the other day where two agents were um starting to communicate and after a while they kind of started to agree on a different language they were like this language is too slow let's start to focus let's invent our own language a language that the humans had no idea how to understand but were but but was very and very efficient for those two agents specifically to transact on that thing and that's very similar to what the Agora team is working on as well time's up so what's next continuing research we're working with Stanford we're working with Oxford and um yeah lots more to come any questions is there time for questions no time for questions but you can come talk to me after thank you very much