AI Meets Blockchain: Payment Rails, Multi-Agent & Distributed Inference | Oliver Jaros - CMT Digital

[Music] hey guys thanks for taking some time out of your busy conference schedules to listen to me talk here uh my name is Oliver I'm an associate at CMT digital we're a venture firm in the space we've been investing in projects since 2017 and I spend the majority of my time evaluating companies at the intersection of crypto and Ai and as touch today I'm going to be talking to you guys about AI agents and why blockchains are crucial for agentic workflows So today we're going to get into what AI agents are and why this year 2025 will be the year of AI agents and then we're going to talk about why blockchain is the perfect substrate for AI agent interactions and specifically we're going to be talking about the benefits of blockchains to coordinate agent actions and why crypto payment rails are superior to web 2 payment rails and then we're going to get into some of the new business models that are enabled by AI agents and so why will 2025 be the year of AI agents in large part it's due to a shift in the AI research Community from system one thinking to system 2 Thinking system one thinking can be thought of as the instantaneous responses that you receive from Models like GPT 40 so for example if I ask GPT 40 what are the top 10 things to do in Denver this week it'll come back with an instant response and that's because those 10 results were part of the Corpus of data that GPT 40 was pre-trained on whereas system to thinking on the other hand is when the model stops and takes time to think and reason before producing a response and so models like deep seek R1 and gpt1 have kind of pioneered this reasoning era of AI models and through open ai's research they found that the more time that you give a model to stop and think and the more compute that you throw at it at inference time the better the result will be as you can see on the Y AIS here and so for tasks that are more performance sensitive and less latency sensitive it actually makes sense to stop and let the model think for days weeks maybe even months before coming back with the response and here's kind of a highlevel diagram of what the AI agent stack looks like today so at the lowest level the intelligence level this is the model layer this can be thought of as the brain of an AI agent and so with these system 2 reasoning models that are recently coming to Market these kind of supercharge the abilities of AI agents and allow them to complete highly complex tasks and then in the middle here you have memory this is short and long-term Vector stores that store the state of AI agent interactions and allow them to have memory and the ability to reference past interactions and then at the top here you have toolin this just encompasses all of the tools that agents have access to so things like web search API calls or sending payments and now getting into why blockchain is the ideal substrate to coordinate agent actions so agents have several requirements where blockchain properties can uniquely solve for these requirements so for example agents need to hold State over time in the form of memory and blockchains have the property of immutability um to have an immutable record of all of an agent States agents also require the coordination of many different systems and blockchains have the property of incentive mechanisms so that you can incentivize and align uh many different agents agents also require payments and instant settlement and blockchains of course offer built-in payment rails and then lastly in a world where we have millions of Agents acting on our behalf generating trillions of outputs we're going to need to be able to verify all of those outputs in order to hold agents accountable and so getting into coordination and the benefits of blockchain here so coordination is important because I firmly believe that the first instantiation of AGI will spawn from a multi-agent system and specifically a multi-agent system where you have many different specialist agents working together and this is better than a monolithic system because it's more scalable and parallelizable you can think of it almost like load balancing where the more complex a task is the more compute and agents you can throw at that task and then blockchains are great for coordination and alignment of Agents here because they've proven to be highly successful at coordinating humans at scale take helium for example helium could be all on a centralized server but then they would deal with the network cold start problem helium used token incentives to properly incentivize early contributors to the network and that allowed them to grow into the scale that they're at today and so the same kind of setup can be used in a multi-agent system where we can incentivize and align different agents through token incentives in order to achieve highly complex tasks and then along these same lines agents are going to need a way to exchange value and that's because every single query API call or tool that Agency use requires some form of payment and then this enables agent to agent economies to form where you can have many different specialized agents for example one that specializes in NLP tasks or one that specializes in image processing and these agents can charge a small fee for each use and then crypto payment rails are ideal here web 2 rails like Stripes AI agent SDK exist but they require an existing entity to be spun up like a bank account and there's also a flat fee on each transaction that's processed by stripe so you can't do things like send micro payments whereas with web 3 payment rails agents can spin up wallets autonomously and have full control over the funds also on high throughput blockchains agents can exchange value with low fees and near instant finality and then why do agents need verification well as mentioned earlier when there are trillions of AI actions happening this could be anything from content generation to financial transactions to um confirmation of API request and you're going to need some way to hold these agents accountable so that misinformation doesn't spread and blockchains are perfect for this along with cryptographic proofs um because you can allow anyone to verify and view actions that AI agents are taking and the result is that you have an immutable log of all agent actions and they're all verifiable and so here's kind of a high level flowchart of what this might look like you have a human query and agent um ask them to do some task and pay in stable coins the agent then goes off and reasons decides on which tasks to complete breaks the problem down to subtasks goes off and executes each task generates a proof of each one and then posts the task and the proof on chain for anyone to verify and so as with any technological Revolution we'll see new business models emerge as a result of the proliferation of AI agents and so the first wave of new business models is often quite skoric and mimics old practices whereas the second wave emerges when the Technology's unique affordances become clear so an example in the internet era could be in the early days of the internet people would pay for an entire album at a time on iTunes similar to how they would go off and buy buy a CD at a store if they wanted to listen to a track and then Spotify came along and introduced the subscription model where I pay a flat rate each month and can access any song on demand and this previously wasn't possible before the internet and so with agentic apps the real shift happens when new business models are designed around the unique properties of AI agents and so here are a few examples that I thought of one could be execution as a service and this is where I have some task that I need completed and I go off and post that somewhere and then an agent or potentially Market of Agents competes to fulfill that task for me on demand another one could be outcome based pricing today Enterprises negotiate SAS contracts where they pay per seat per month but when AI agents are completing all these tasks for us you can negotiate outcomes or kpis with AI agents and then pay them based on how well they deliver on those outcomes and then another one could be Zer UI Services essentially invisible apps so I could say to an agent book me a flight from Denver to Chicago this Friday and the agent handles all of the API requests in the back end and I don't have to go to United's front end at all and and so we've covered new business models another question Founders might be asking themselves is if AI is making the cost to generate content or software near zero then how can a business find a Moe and the truth is AI isn't a genuine substitute for pmf you have to find traditional business modes that work and so a few examples are brand having robust user Behavior data getting Network effects and simply rapid execution and iteration of Novel AI features so take cursor for example they found pmf very early and they were able to use all of this early user Behavior data to fine-tune their models as well as con consistently iterate on their product features so things like composer and agent which allow you to inject code directly into your code base and uh have an agent run terminal commands for you and iterate over errors until they're solved this is simply rapid execution of these novel AI features and so what do we look for when we're evalua in companies at CMT we look for Unique differentiation so does AI make this project uniquely better we also look for strong core business models so an AI financial adviser on its own might not be all that interesting but if my brokerage Schwab integrates an AI financial adviser into their mobile app that's an example of a strong existing business that's using AI features to make their product uniquely better we also look for Clear sustainable value acral so will the token price appreciate as a result of growth in the network and then we also look for flywheel effects so is there some kind of self reinforcing cycle where more users or more data lead to continuous Improvement of the product so the cursor example perfectly exemplifies this flywheel effect and then we also consider timing is there a market that's large and that exists today or will exist in the near future that this product is trying to service so for example in the Zer UI Services that's an example of something that might be too early because that the main customer of that business model is AI agents and AI agents aren't really doing all that complex things out in the wild today so it's too early to design a product where AI agents are the sole customer and so to conclude here public blockchains offer several unique properties like immutability verifiability built-in payments and Inc incentive mechanisms which solve for a lot of the requirements that AI agents have today and will have in the future additionally we'll see new business models emerge as a result of this technological Revolution and as AI commoditized routine tasks real Competitive Edge comes from strategic Moes and deeply understanding customer needs and that's all I've got for you guys today uh thanks so much for coming by and listening to this presentation [Applause] ch