ASI Workshop I Rishank Jhavar

[Music] Hello and welcome to the ASI workshop organized by artificial super intelligence alliance. I am Rishang Javar your host for today and I lead product marketing at fetch.ai AI which is the part of ASI. Today we'll have a look at search and discovery of AI agents. We'll walk through how to build an AI agent, how to host an AI agent.

We'll learn a lot of things about communication of AI agents with each other. And we'll also see how an AI agent could be deployed to one of the world's first agentic LLMs which is ASI1. This workshop is organized by innovation lab by fetch.ai for it global New Delhi and I am personally looking forward to see you all at the hackathon itself. Me and my team will be there at the hackathon and uh we will be assisting you with a lot of things at the place.

Just to give you some context about this workshop, we would be walking through a general structure of how to build an agent, how to host an agent, how to deploy an agent and we'll also learn some of the agentic communication methods. We will have a follow-up detailed workshop at the venue itself where you will learn more about the broader stack of ASI. So let's jump into the overview of the workshop. I guess the very first question is what is fetch.ai which is something that we will see and we'll answer.

We'll then move on to some of the tool stack, something that you will be working very closely with at the hackathon. We'll then move on to agent search and discovery, which is the key element of fetch.ai. And right there, we'll start building our first AI agent. And we'll see the communication processes of AI agents.

We'll also explore some of the development approaches. And then we'll move on to two of my favorite topics which is chat protocol and ASI1 LLM. Um basically how to send your agent so that the entire world can use it on ASI1. We'll also have a look at some resources and then read me and ranking tricks. Basically some of these secret sources that can help you win the coveted ASI prize.

So let's just jump into it. I guess the first question is what is fetch.ai? So fetch.ai is part of ASI alliance an alliance that is working towards decentralizing artificial intelligence and that is where fetch.

ai techstack offers an entire ecosystem for developing all your AI powered services. So you build, discover and transact using a single platform. Fetch.AI is enabling the agentic economy by using a very unique search and discovery method for AI agents. You can basically break fetch.

AI down into three components. One is ASI1, the consumerfacing personal AI agentic LLM that anyone across the world can use. Two is Agent Worse, an amazing amazing development platform where you can build and connect your AI agents. It's equivalent of Versel for React.js.

You do not really have to do a lot to create an agent. Just FYI, I'm a product marketing guy and I can easily create and deploy an agent using fetch.I stacket. It is that simple. And then we have the core opensource library which is u agents which is something that we will also learn about and you'll have a lot of resources learning about u aents.

So this is the broader tool stack of the fetch.AI ecosystem. We'll just quickly walk through pretty much everything. One is build. You can build an AI agent using U agent or using other frameworks.

It doesn't matter. Uagent is one of the important AI agent frameworks out there. But if you're familiar with Langchain, Crew AI, Autogen, you can still build with fetch.AI stack. It is a completely universal plug-and-play model that encompasses a wide range of AI agents in it.

With U agents, you can build using U agent. You can directly push it to ASI1 LLM and you can use chat protocol which is a very simple protocol for your agent that enables it to chat. You can also use other frameworks like Langchain, Crew AI, Autogen, you name it. We have all the major agents and agentic frameworks that you can use via ASIM. If you're building using U aents, you can use agent war.

You can use agents mailbox to host local agents. Something that I'll briefly talk about when we move on to the agent topic. And if you're building using other frameworks, we have very easy u agent adapters out there. Um we have a very amazing guide with innovation lab that I'll share the link to as well. so that you guys can use it.

And the third step is to deploy. You can deploy an agent on agent wars or on your own virtual machine. The advantage with deploying agent wars or deploying on agent was are multiple. It's it's absolutely crazy how easy it is to deploy an agent on agents which is something that we'll also see. If you're hosting your own agent on your local machine or virtual machine, you can do that as well.

It is completely infrastructure agnostic. And finally, you can test your agent. You can do that directly on agent words using chat with agent. You can do it on agent words or if you have a UI created, you can also test it locally. If not UI, you can also test it on CLI.

And finally, and the most important step is discover. This is something that is amazing, unprecedented and no other framework has ever attempted before. You can discover your agent and the entire world can discover your agent directly on the ASILLM. Um, granted, it should be an amazing agent and something that they really find useful. So, this is the broader tool stack.

We will be diving deep into creating an agent using U agents. We'll also see how to add chat protocol to it. We'll connect it via agent wars. We'll deploy it to agent wars. We'll test using agents and chat with agent.

And then we'll also see how to discover your agent via ASI1. Um the most fun step I would say. So this is the broader overview of agent search and discovery. You have your UI AI agent. You can build it using u agents.

You can build it using lang. You can build it using crew. You can build it using autogen. You can build it with pretty much any framework that you want. The U agents register on the Almanac contract which is a smart contract.

It makes them listed and searchable in the agent marketplace. Agentw marketplace is a huge public directory for discovery which is something that we will see. And then ASI1 can automatically search agent words ranks the candidate agents and orchestrates the best one for a user school. So for instance as a consumer if I want to see or if I want to connect with an agent that can create an anime related picture of me I'll just go to ASI I'll prompt it to create an anime picture of me. It will do an agent search and discovery, find the right agent on agentworth.

ai and it might discover your AI agent if it's the best one for my goal. It is as simple and this is the most fun part. Let's build our first U agent. So let's see this step which is the agent creation. To do that you simply need to navigate to agents.

I have already logged in via fresh account. So you can see what you will basically see. You will land on the marketplace of agent wars. U marketplace has bazillions of agents. By the way, there are recommended agents which are topmost agents that have been recommended by the community.

Um you can go ahead and you can explore any agent yourself uh basically on agent words. Um because it's it's super simple and super straightforward. So for instance, this is the Google API geoloccation agent. It is a very simple agent that returns the geo code of a given address. You will see a very nice readme.

md file. Here you'll see the rating of that agent. Uh this is based on the complete SEO of the agent and you will also see the all-time interactions of that agent in a general sense. Um there are some protocols that will be associated with every agent. we'll find that any agent that can be connected via ASI1 has this important protocol uh and I would say the most important protocol which is the agent chat protocol um which enables people to interact with this agent on ASI1 and this is the general overview of the marketplace as a whole.

I guess uh this would be a good point to kind of start creating our first AI agent. To do that, navigate to the agents tab. As you can see, this is a fresh account. I do not have any agent yet. So, I'll just create an agent for now.

I'll just create a very simple agent to understand a little more about the agent ecosystem. I'll click here. I will basically type Rashank's first agent. I'm Rashank. And I'll create an agent.

It is that simple. Once you do, you will land on this page which will give you a general idea on how to get your agent found in five simple steps. How to add a chat protocol, how to create a readme, how to choose your agent handle, um how to start your agent, and how to receive an interaction. Now, this is not exactly an agent that I would like to get discovered per se. So, what I'll do here is that I'll just navigate through a few things and I'll showcase uh what an agent would look like.

um if you sort of start building it. So this would be the overview of the page. Something similar to that that we we just saw in the goo Google geoloccation page. Um it says my first agent which is the name of the agent. A simple example designed by me.

Um you know this covers creating an agent instance. This uses the on interval decorator to run the task periodically and it also accesses the agents context to log the messages of this agent. Now I'll head on to the build tab and I'll see an IDE right here inside agent works. Now you can see a few comments on there and you can see the structure of how an agent looks how how an agent basically looks like you know um basically you can just go from u agents import agent and context uh you can see the onvent here you can see that we are logging hello I'm an agent and my address is ctx.agent agent address.

By the way, this is the exact address that you will see when people discover or want to discover your agent in a general sense. So, this is going to be your agent's address and you know the agent can basically run um just to showcase I'll start this agent. You can see that there are a lot of timestamps here and yeah hello I'm an agent and my address is agent 1q1 blah blah blah blah blah and this is your uh AI agent which has successfully started and by the way um already cued for almanac contract registration. So creating your first agent is as simple as that. You just need a few clicks.

You can also modify this agent completely as much as you want. I'm not doing it right now because there is one more important concept that we need to see which is agent communication and that is where we'll basically navigate to I'll stop this agent for now because I don't really need this one. I'll hit back. I have my first agent here. By the way, all your agents on agent will be listed here.

I'll launch another agent. It asks me if I want to use a chat protocol agent and if I want to create select an communication type. But for now, I'll just create an simple agent was hosted agent. I'll click on agent communication. It will create two agents Alice and Bob and demonstrate how to send messages between them.

So I'll basically type my name here first and then I don't know I'll interact with Alice. I have just created an agent and you can see that there are two agents here. I'll just kickstart Rishank. So Rashank or Alice is a message sending agent which initiates communication by sending a message to another agent which is the Bob agent. The other one that we just created.

Um we have a few concepts that are included here. The most important one is using the oncore message decorator to handle incoming message which will be evident to you once you see the code if you're already a developer. And I'll just head on to build. So as you can see what this agent will do is that it will use an onssage decorator. It will log the receive message along with the sender and it will say received message from sender which is msg.

m message. Um you can see that this is the message right here which is hello there Bob and we'll also say that message has been sent to Bob in a general sense. Um just before starting this let me duplicate this and I will also open the second agent that we created before which is the Alice agent which is inappropriately named Bob because I actually changed the name but I'll simply rename it to Bob. Super simple right there. And I will just start this agent first.

Uh you can see here by the way that uh this agent uh will basically just receive it. Um if the sender is the previous agent that we just created, right? Okay. Those the agent has started. It has been registered to Almanac API fasttrack.

And now we will head back to the rishank agent and I'll start this agent. So it will start the agent just about to send a message to Bob. Um message has been sent to Bob. The envelope has been dispatched to the hosted agent which is the other agent that we basically wrote about and it has successfully started the agent. The almanac contract registration is skewed for processing and we have received the message from that agent which is hello there Alice because Bob here has received message hello there Bob.

As you can see agent communication via on message is super super simple. Two agents can autonomously communicate with each other in a very simple way and it is pretty straightforward. right here. You can try this demo out yourself. I will highly recommend that you do because this will teach you a lot of concepts regarding agent communication in a general sense.

I'll stop both of these agents because we don't need them and I will head back to the most important and the most fun part of today's tutorial which is the chat protocol using LLM. Now I will start by launching an agent. I'll not click this for now. You can also click this because this will basically give you multiple communication methods in a general sense. But I'll go the classic way and create an agent versus hosted agent.

For now I'll pick this chat protocol using LLM which creates a simple AI assistant. By the way, this uses um a certain API another LLM just to showcase how it looks like. So we'll be using open AI um and I'll just create a very simple AI assistant that is an expert in a specified subject. So for now let me call that agent um I don't know um maybe bird's agent. So this will create my agent automatically.

We'll just head over to code which is the most important one here. Um yeah so uh by default we are using ASI1's LLM endpoint and model. Uh we can insert an ASI1 API key by create an account at ASI1.ai/dashboard/ API keys which is something that I'll show. We'll create a new protocol which is compatible with the chat protocol spec.

Now this ensures compatibility between the agents. And then we'll define the handler for the chat messages that gets sent to your agent using the ctx. Method. Uh we'll send the acknowledgement for receiving the message. We'll collect all the text chunks.

We will also basically add some responses like you know I'm afraid something went wrong if something actually goes wrong. Um and some of the most one of the most important point here is that we will be talking about a specific subject matter here which is something that is sort of required. So let me just go back and I think that by default it is set as birds um which is something that I actually wanted to add in a general sense and this will be set as default. You can go back um by the way you can also add a chat protocol to the agent. You can do several things.

Uh once you start the agent, chat protocol will automatically be added to it. Um I'll just quickly modify this readme file. I'll call it bird's agent. This agent will give you a lot of information about the birds. And I will just release this um everything there's to know about birds.

This agent is an expert on birds. Guys, it is very important to have a very good readme about this. This is not a good readme, by the way, but I'm just doing this for the demonstration purposes. It is super super important that you write a very detailed readme here so that your agent gets discovered properly via agent words. Right now location is not exactly needed for bird's agent.

We have already walked through the code of this specific agent and now we'll just start the agent. All right. As you can see the system has started the agent. The system has connected to Almanac API fasttrack. The almanac contract registration is skewed for processing and the agent has started.

So we can see that the chat protocol is added to the agent via the code. We have a readme. Uh we can choose an agent handle later and obviously uh we can trigger this agent and receive an interaction. Um you know once once it gets found we'll know it right. So I can either just go to ASI1 and basically query this agent.

But before that, I'll need to add an API key, right? Because it doesn't doesn't really have an API key here. So, in order to do that, I will navigate to this link. I will create a new key. I will say birds um project maybe and then it gets created.

The key has been successfully created. I will copy this key. Please note that you will have to copy this key otherwise you'll not be able to see this again. Um we are doing this for security purposes. I'll just press done.

Uh I know that you might try to use my API key as well but I'll try to delete this so that you can use yours. Uh there's a certain kota that we assigned to every API key. I will need to stop this agent first because it's already running uh and it's in read only mode. Then I will insert my API key which is something that I took from the next step. And then I will start the agent again.

Now the agent was shut down. It is starting again. It has an API key and the agent has completely started. Or I can directly head over to ASI1 right here. Uh which is my AI and it will answer me pretty much everything that I want to want it to answer.

Uh, I'll click on agent search in order to search for an agent. I've already copied the agent which we just created in this step and I'll ask it what is a peacock because I have a lot of peacocks in my region, right? So, it's it's natural. I would like to know more about it. The ASI one will go back.

As you can see, it has sent the message to the bird's agent agent. Um, sorry for the redundancy here. I should have named my agent better. Um, now once the agent replies, the response will be sent to you and the bird's agent has responded. A peacock is a bird from the fezen family scientifically known as Pavu Chrisatus.

I guess the term peacock specifically refers to the male of the species blah blah blah. Now I know everything about peacock and that's the fun when you work with an agent. Uh, I can click on the agent and I can also see the details of the agent. I'm seeing all this right here because I have just created this agent myself, right? Uh but you might not see because this is a general marketplace agent in a general sense.

And that's how you can quickly build an agent and get it discovered directly on ASI1. You can also chat with your agent right here on agentw.ai. Um it will give you similar responses. You can test this agent right here.

um and it should give you similar responses as um how it is giving on ASI1 directly but I'll highly recommend that you test it directly on ASI1 because I don't know it looks cool. Uh so that's chat protocol using LLM. Super simple, super easy and super intuitive to build. Also now we'll discuss about the three different development approaches that we have. One is a hosted U agent.

Now the best part about hosted U agent is that you do not have to worry about anything. Everything will be taken care of by agent. You do not have to worry about infrastructure at all. So you do not need a DevOps engineer in your hackathon team. You can just do it yourself.

Um it is pretty straightforward. The only caveat with that is that since uh this is a public marketplace, we have to be very careful about security and that is why we usually put uh only certain Python packages here. Um we put a lot of them by the way but uh you know we we sort of have to kind of restrict those packages in a certain sense. Uh you'll find all the major packages here. There's a lot of uh built-in uh Python library that we already have.

But obviously it's it's sort of limited in a general sense. If there are very specific Python packages that you want to integrate, you might want to create a local U agent. I'll highly recommend that you don't do it. I understand that it gives you more control over your agent, but then deployment becomes a hassle and you have to use a VM and everything. Uh with a hosted U agent, you can directly just put it on agent wars.

Um and then obviously if you're using a local U agent, you can directly connect local agents to all agents on agent wars. Uh which is quite simple, quite straightforward and quite intuitive as well. Uh we have a very detailed guide that you can basically check out. Uh it's available on innovationlab uh.fetch.

ai. I'll also be talking about this resource at the end but just giving you a briefer here. So this is something um and then we'll see how the magic works how chat protocol works generally. So chat protocol is a fetch.ai's proprietary protocol that generally helps your agent interact and get discovered on agents itself.

It features natural language processing. It makes agents ASI1 compatible. It has an acknowledgement reliability and it has session and streaming control. Something that is super important. Here you can see um how everything works in a general sense.

It is very evident from this diagram on how the chat protocol works. So say that you have created a weather agent and that agent is hosted on agent wars. Now what it's going to do here is that uh once a user queries ASI1 it will send a chat message then it will extract the parameters from the query. For now let's say that it uses an open AI agent. It returns the extracted parameters in a pedantic model which would be the structured output response something that we just saw in the code that we showed earlier.

And then it sends a response to the ASI1 LLM. ASI1 recognizes the message and uses the handler chat acknowledgement. Now you'll be wondering why we are only using U agents. You can also use other frameworks per se. So we have a very quick guide that you can use to register a lang chain or a crew AI agent on agent.

Uh let me just quickly showcase you the guide that we have put together. You can obviously head to innovation lab talks. By the way, this is an insider tip. These talks are 100% designed for hackathon participants so that you can get started quickly. Uh it's always super easy to just hit the innovation lab.

fetch.ai/resources URL and see all the guides here if you want to win this hackathon. It's it's pretty straightforward in that sense. Um but yeah, these are the U agents adapter. They connect entire ecosystem of all the AI agent frameworks.

Right now we have picked three of them for these adapters and three of them are directly available. One is the lang chain adapter that can connect langchain agents. Two is the lang graph adapter that can directly connect lang graph. And then we also have the crew AI adapter that can expose crew AI's collaborative agent teams as U agent so that you can use them directly. Right?

We'll just see a very quick diagram on how to use it. First is that you can choose your track. If you want to use fetch.aiu agent framework, you can directly use u aents which would be directly connected to agents and you can directly access it via custom application. You can create a diagram for that.

Um sorry you can create a UI for that and then it will also be directly connected to ASI1 uh where you can access it. Otherwise if you want to use lang chain crew AI or any other agentic framework you can connect it using u agent adapters um that are available right here. You will find a link to this as well. Um so feel free to use that and you can just create um your in agent and then submit it to the fetch.I track.

So that is something that you can do. Now we will move on to the next part which is giving you some idea about the ASI1 agentic LLM itself. So first off let's see what an agentic LLM is. This is the ASI1 official documentation. Another great resource that you can use for the hackathon itself.

ASI1 has an agent tech LLM and it has several models. Let me just jump onto the overview here first so that I can show you the models. Um, we have ASI1 mini that balances performance as well as speed. So, if you want to create a generic application, this is the go-to model for you. I mean, you should simply use this directly.

If you want something that just gives you very quick response like a response in a microcond, then just use ASIS. It's a it's a no-brainer to use it there. If you want to do or perform some very complex task then you can move on to ASI1 extended which has enhanced capabilities for all the complex tasks. ASI1 extended is specifically used for some complex task or some complex workflows. And then you have ASI1 agentic which is highly optimized for agent interactions.

Um if you are not really that keen on using the core LLM, you can also use ASI1 agentic for agent interactions. If your application is mostly agent based and then finally if you have a application that is very heavy on generating graphs or data analytics then we have ASI1 graph for that. Uh for now let's head on back to building with ASI1 which is the agentic LLM. Um you can build applications directly using the agent marketplace as we just saw and they would be discoverable via ASI1. Um you can just use the quick start application if you sort of want to create your own agent.

Um you can obviously directly do it on agent but if you want to do it locally you can use uh a quick start guide for that that is present right here. Um and you'll also find a few things in the same guide. Uh that would be session management because obviously the models will require some session persistence to maintain the context. Uh then we have the async agent processing. um agents from the marketplace might need some time to complete the task and that's that's where you can use the async processing.

Uh you can find a few examples here as well. And then something that is quite important is that you try to use the best practices for the proper search and discovery of your agent and you can do it right here. So that's that's pretty straightforward. Another guide is for tool calling which is something that you will obviously do. This enables your interaction or integration with any API tools or your own code to retrieve live data, perform task or trigger actions based on user input.

Right? So it in an archelit enables you to integrate your custom code with the chat completion API in ASI1. Um so when you give access to defined tools your models whatever you pick can choose to call them based on the conversation context. Right now it is enabled with three models which is ASI1 mini, ASI1 fast and ASI1 extended. You can see a very simple example here.

Um feel free to check it out and you will also see an example response here when you make the tool call request. Uh you can see several sample tools um that you can try and yeah uh you can also head on to tool definition. You can see additional configuration. Basically this guide will end to end provide you everything um for everything that you need for tool calling in a general sense. Let's also quickly have a look at structured data to generate structured JSON style outputs by prompting models to generate specific fields or return specific fields.

uh you can define your response format and you can get example output in that format. Uh please note that this model is instructed to return JSON only and uh you should try to enforce all the schema server side and reject the invalidation. That's that's just for general good practices. Uh so that would be structured data and obviously we have made everything a open AAI compatible in a general sense because uh we sort of recognize that it is one of the most popular LLMs out there and you might need to use its chat completions API. So that is why ASI1's API is 100% compatible with OpenAI's chat completions API format.

Uh so this means that you can use all the existing OpenAI client libraries. you just need to change the base URL to start using ASI1's agentic models right so that is going to be uh the core change here but you know you can just use it directly right here you know so we have a very straightforward flow for that I just showed you an example that kind of uses this um so that might be very useful for you um and as I mentioned before you need to choose the right ASIN model which is based on your use case for this We also have async communication that we just saw by the way in the open guides and we just sort of read a bit about the multiple models earlier. So that is something that you can also do. Now the best part about building with fetch.ai is the enormous amount of resources you will have at your disposal.

First off is the ASI1 documentation that we just saw. Let me close the other tabs so that you can see everything clearly. This is the ASI1 developer platform. It has everything right here including a tutorial for chat protocol. Then you have the agent documentation which is how to do a proper search engine optimization of your agent which is super super important by the way.

Um I feel like this is one biggest thing that you should follow in this hackathon is to properly search optimize your LLM. One of the key metrics that we would be judging on and this is an insider information if you made it uh to this part of the video is that we would be searching your agent on agent wars via the query that you give us and agent war or or or your ASI1 LLM should find that. So that's that's something that we would be doing. So yeah um so yeah you can go through the entire guide. Um it's going to take another session if I basically talk about SEO but it's a 15minute guide.

You can just quickly read this one just to understand everything. Uh I'll come to this link later and at the last but you know we also have the uh documentation uh that will be a quick start guide for the u frameworks um adapters as I mentioned we have innovation lab documents here. We also have the chat protocol guide here. This is the diagram that I just showed you for the weather agent example and uh you can also try this out directly. Um we have MCP server integration on agents innovation lab examples and front-end integration with U aents.

These three links and all of the other links most of the other links are directly connected to this. This is going to be your onetop for hacking hackathon. So this is basically the go-to guide for every hackathon participant that we usually ask to follow. This is the secret sauce ladies and gentlemen. If you haven't really gone through this video as well and if you have somehow fast forwarded this video and if you landed on this section the only one recommendation that I would like to give you is to go to this resource.

It has everything that you need everything that you need to build an agent. So you will be able to understand AI agents you'll be able to understand agent creation the adapters agent to agent communication chat protocol what is transact which is agent to agent payments. If you want to enable payments in your application, everything there is to know about agent wars, everything there is to know about ASI one um including image generation, chat completion and how to integrate MMCP and this is something that you might need because you know it's quite simple and quite easy. And last but not the least are the examples. We have so many examples here.

so many examples that you can just go through and understand everything yourself. Um, something that I find really useful are the innovation lab GitHub examples because we have a direct repository. By the way, if you're on here, it would really help us if you sort of start this repo. But we have the innovation lab examples repo that have directly put a lot of examples on here that you can read directly. Also by the way if you want to really sort of uh do a cheat code in this hackathon we have fetch development rules for cursor IDE as well.

Um you know just something to think about. Uh by the way I'm winking I'm not on the screen but yeah uh this is something that we have here as well. U and we have a lot of other examples you know Nike product agent blood report analysis agent and I don't know an overwhelming amount of examples right here. So yeah guys you can check out resources. I would highly recommend that you do because it's quite quite good and obviously feel free to ask and reach out to us.

We will be present on the ground at the hackathon. I will be there. Dave Johar and Gotham from my team would be there. Um Churag our senior developer from Metallics and Fetch.AI he would be there as well as uh um one more developer would be there.

So there will be several developers directly. The innovation lab team will be on the ground. the ASI team will be on the ground to help you guys out and build with the fetch.ai and ASI stack. So this was today's presentation.

If you have any queries, any questions, feel free to reach out to us. We have a very active Discord. We have super active GitHub as well. You can raise issues directly there. Um or you can also email me directly on r i sh.

j at the ratefetch.ai. that is rishang.chavarfetch.ai and we will try to answer any questions that you have.

So that's it guys. Uh I wish everyone who's watching this video all the very best for the hackathon and I hope the best team wins. Thank you. [Music]