A2A Protocol Workshop: Build Interoperable Multi-Agent Systems | Community Webinar

Key Takeaways

All right. Hello everyone. I hope you guys are all doing well. I welcome you all to this webinar where we are going to discuss the A2A protocol. We are going to do a uh live workshop and we will be building an interoperable multi-agent systems. So this webinar is a a series that we are actually continuing from what we did in our last two webinars about agentic AI protocols with MCP integrations in which we interacted with 7,000 applications with uh seamless MCP client and MCP servers and we talked about uh agenti protocols A2A and ACP. In the coming webinars we'll be talking about the rest of these topics. So uh today's agenda would be mostly around A2A protocol workshop and you know me I've uh from my previous webinars as well I am Zad Ahmed I've been working here at data science dojo for more than two and a half year now as a senior data scientist. So if you guys have any questions you guys can put it in the chat and let me know where you guys are joining from. Um uh you guys can put your location in the chat and then I'll continue uh watching over the uh chat thread. So if you have any questions just put it there. All right. So getting started with today's agenda. Um in this protocol A2A protocol workshop we're going to first discuss the key challenges in agent development which actually started in like 2024 2025 people started building uh AI applications on LLMs and uh after GPT3 was released people started using the APIs to make something similar to chat GBD. So since then there have been multiple challenges uh while developing these agentic AI applications and those challenges have been resolved by using agentic AI protocols. We started with web applications software development we used to use some protocols like REST APIs, HTTP protocols, all of these things. Nowadays even LLMs and agentic AI development has protocols. So these protocols are mainly uh the most viral ones are MCP, ACP and A2A protocol. So we have already discussed about model context protocol which is MCP. I'm going to go over a quick recap of our previous sessions as well just so that we are connected and then we're going to continue our discussion on how A2A agent to agent protocol is different from MCP. We're going to discuss a simplified interaction flow with A2A. How A2A works. We're going to discuss core concepts of A2A. Um what are the specific uh specific details of these methodologies, how to interact with multi-agent orchestrations. Then we're going to discuss some powerful use cases across different industries where A2A can be used. And at the end of uh this session, we'll also do an hands-on exercise building a multi- aent orchestration with A2A protocol. So if that sounds interesting, let's quickly dive into this into today's topic where we're going to discuss the key challenges in agent development. All right. So after uh the advancements of uh lang chain lang graph u I'm I'm probably sure that you guys have used most of these u uh libraries to build a applications on top of llms whether it could be crew crew ai or lang or llama index or something along those lines every other uh every other library has some challenges while you're devel developing these agents. Uh one of those challenges is that u uh these multi- aent systems have orchestration problems that coordinating between multiple agents to perform a complex complex task requires clear workflows, data passing and error handling across agents which lead us to write a lot of a lot of code uh in order to orchestrate uh the uh between the multiple agents uh in a particular query. to your task. So, orchestration was a problem. Then building these robust agents, it requires managing complex uh complex workflows and consistent logic and behavior. So, there was some gap there as well. And then interoperability was another issue that how do these agents with different tools. Maybe you want one of your agent to talk to your CRM. You want one of your agent to talk to Slack. You want one of your agent to talk to your native database or something along those lines. Interoperability was was a big challenge. I remember when uh in in 2024 2025 when we used to build integrations it used to take us uh a couple of weeks to build um integrations with teams to build integrations with Outlook and Slack around those using their APIs. But nowadays with agentic AI protocols and namely model context protocol MCPS this interoperability has become so much easier that with just few clicks you can deploy an MCP server and use your MCP client to interact with it. So these are the challenges that have been resolved by these agentic AI protocols. Then there is control and safety that authentication the tool the LLM is using tool on its own there should be some guard rails and evaluation associated with that. So we're going to discuss those things as well how these challenges have been resolved by agentic AI protocols. So in agentic AI protocol the most viral one was MCP model context protocol. So let's do a quick recap on our previous webinar. Um I'm going to request Lisba to share the links of previous webinar recording as well in the chat. So if you guys want to go and do a recap before watching this webinar, you guys can do it like this is a recorded one. So you can uh watch them afterwards as well. So model context protocol in in u so one example I always give to my examples when I'm uh uh to my students when I'm teaching them about model context protocol is in uh traditional uh programming we used uh every software developer used to build rest apis so that human can interact with uh human can interact with their software but nowadays human are rarely interacting with the software but nowadays LLM have started interacting with these APIs and software. So, Anthropic went ahead and created this model context protocol which is which is more like a wrapper on your API that wraps your API in such a way or wraps your API in such a way that it creates tools out of it which can be easily used by the by [clears throat] these LLMs. So nowadays if you are a software developer you are working on an ERP system or any software and if you want to use if you want your LLMs to use your APIs you have to create an MCP server or a tool wrapper on top of it and that's the brief introduction about model context protocol and if we dig deep into this uh this MCP architecture starts with we have a MCP client that actually u uh connect us with the server. If you have worked with databases, we used to create client in order to interact with the database in order to do our CRUD operations. We're doing the same. There is a client that is responsible for interacting with the server to do the CRUD operations and all of the business logic lies in the MCP servers itself. So the client is just there to interact with how many tools this server has and based on the user query. Uh it will fire those tools and get the response back to the user writing as a final answer. So um and if you want more details about MCPS uh you have uh a webinar on we have done an webinar on agentic AI integration uh with MCP protocol. You can watch this and then there is a number of hands-on exercise we did. We created a Zapier MCP client and then you are able to interact with 7,000 applications with MCBS. So that's already there. That was a quick recap on what we did in our previous webinars. Now moving towards agent to agent protocol. So when people started building applications there used to be a chaos around this that how to navigate multi- aent collaboration how to do it correctly with minimum effort. So uh there used to be a problem that uh companies started building agents that works really well on some environments. Let's say uh I created an HR agent for uh for my HR department on Langraph and I created a marketing analyst on uh Google with Google ADK agent and then I created a sales assistant with Creai agent. uh prior to agent to agent protocol there was no such uh no such uh no such solution to actually interact with all three of these agents in one with a one host agent. So we used to actually run uh this agent separately, this agent separately and this agent separately. So this has been a a a a very big problem for companies who want to interact with multi- aent uh systems. Let's say if you are the CEO of the company and you want all the details from HR, your marketing assistant and your sales analyst, you want to have one agent that can actually orchestrate the your work that can actually divide your work and can get information out from all three of these agent. That's where it all started that agent to agent protocol is um is created in such a way that you have your host agent which is responsible for doing all of your interaction. It is responsible for managing uh all of your conversation. It is responsible for managing the the direction you want your LLM to take. It is also responsible for managing uh which agent to uh to call for a particular task. And this is something we're we're going to uh understand today in depth uh as we move forward. So just a brief uh introduction about why A2A is different from MCP. So MCP is more like agent to tool uh protocol where your agent can actually connect to multiple tools and it actually directs it to different tools. So let's say if this is your agent, it has a agent framework and maybe you build an MCP client within this as well and this MCP client is responsible for talking to the server and then it can interact with different tools. Maybe this could be a CRM, this could be a Slack, this could be uh your HubSpot or something along those lines. So this used to be so you can see there's a clear distinction between agent to agent protocol where it is responsible for uh this protocol is responsible for connecting your main agent your host agent with various other agents that are built on different platforms. This could be built on langraph this could be built on Creai. So, so this give you the ability that you can have uh various text stack of different agents and then A2A protocol is actually responsible for uh your uh agent orchestration. It is responsible for managing the state of your agents. It is responsible for passing the memory and it is responsible for all of the things that are underlying problems for an LLM. So A2A has made our life easier in order uh to actually interact with multi- aent systems. So if anyone has any questions I would be happy to answer. Um um so uh El Musl has so with A2A we might use MCP as well to access different tools. uh 100% correct with A2A uh it it gives more weightage and power to MCP that you you that the agents can interact with MCP servers. So MCB provides uh this the dynamic tool discovery and all of these things and then A2A is responsible for interacting with multiple agents. All right. So if anyone has any questions, I would be happy to answer. Just put uh all of your questions in the chat and then after uh after this thread theory session, we're going to go over uh the queries one by one and before we do our practical exercise. All right. So A2A versus MCP. Um and this has been a question going on in the chat. So the feature is that MCP is um more directed towards u tool discovery and the tool usage and then Google A2A is the the agentto agent protocol which is more directed towards communication between independent agent systems which which means that you can have a crew AI assistant, you can have um a llama index assistant and you can have a Google ADK assistant. And then all of these uh conversations of orchestration has been done through it and whereas MCP is more related to agent to tool interactions. So over here you can see the scope of this is internal that your application talks to uh the MCP servers internally and then uh you can find the number of tools you have access to while uh the A2A concept is more external because your agent systems can talk to external agents as well. These external agents can be on n or they can be on make.com. So maybe if you want to do any integrations uh that you feel that you think that instead of writing the code for it, you can create an anend workflow just go and do it and use a A2A framework to so that you can interact with those agents as well. All right. The purpose of this is to structure context and tools for model processing which mean uh the the purpose for MCP is it to structure the context for and and the tools for your LLM and your agent. While on the other hand uh the purpose for A2A is to enable uh enable external collabor collaboration with external agents. So it allows collaboration collaborating with external agents that are built on different platforms different text stack. All right. If we talk about interaction, so um it is between the application logic and the core language model which is your agent. While on the other hand it is uh the interaction is distinct. Autonomous agent and applications uh they actually interact with each other. And then uh this uh so basically the analogy is the specific grammar and format needed to talk to one expert which means the LLM is our expert and MCP is just um just making the context and tools available for that expert. While the A2A is a universal translator. it can talk to a person who lives in Germany, a a person who lives in Spain, a person who lives in Italy. So it knows uh so it's a diplomatic protocol between nations which means you can interact with various people or having different language sets. So example is this is for uh platform as a service structuring data for its clawed models API call while on the other hand this platform as a service sending a standardized booking request to uh to a different agent. Right? So this is a I would say a brief context about how we started with agentic AI development. what were the challenges in that and how agentic AI protocols have have partially resolved it and now moving forward we're going to understand how A2A works in depth and then we're going to go and build a multi-agent orchestration system as well. So how does A2A actually works with uh multi- aent orchestration that you have uh the client agent finds the agent card of a remote agent for example from a registry or a known URL which means you have uh you have multiple agents deployed maybe some of them are on Creoi some of them are on nitan some of them are built on langraph and deployed on a particular URL so and Then you have a host agent that is uh going to uh on the very first step it is going to discover how many different agents we have and we're going to keep coming back at this example that your host agent is the one who is going to discover. So this is your discovery phase. It is going to discover how many remote agents this u host agent has. Once that is done uh it it does a capability assessment by going through the agent uh description uh the tools this agent have and um the prompt this agent have and the things that is this agent can do. The client agent inspect the agent card. So agent card is uh a way that how we will actually establish uh the connection between our host agent and different other different agents to understand the remote agents capabilities and authentication requirements. So this is your step two. After that in step three it uh the task request uh or the the sending of all of your work. So the client agent construct a JSON RPC JSON uh request that the this is the task according to the A2A specifications. It send this request via HTTP post to the remote agent designated at a particular endpoint and once that is done this request includes task ID method of based on the capability the parameters or any necessary input data. So the query of the user or that has been decomposed into multiple task. So that is being transferred to your remote agent. So, so this is our step three and as we move forward um how do we develop uh these core concepts of agent card that um for every agent for every agent we create um a JSON file for it that this agent has this unique identity. It has these capabilities. This is the end points and maybe if we need authentication in order to talk to this agent because all agents let's say if you if you want to interact with N10 agent or crew.ai agent you need a bearer token or API key in order to interact with it. So, so moving forward to it, all of these agents they have an agent card which is uh a JSON file that in which we uh describe the capabilities of that agent. We describe um the authentication how how you want to authenticate to that agent and all of these agent cards are actually given to this host agent so that it can talk to these agents. So the capabilities, how do you authenticate with it and the tools this agent have and um and the rest of these details like ID of that agent and then your host agent is the one that whenever a user ask a query it goes um the user ask a query your host agent actually uh divide this query into multiple tasks multiple task And based on the availability of remote agents, let's say it has three agents available. So it is going to understand the capabilities of these agents. Based on these capabilities, these multiple tasks would be actually transferred to these agents. Task one, task two, task three, all of these tasks would be transferred to uh these agents. And then once it has gotten the response from there, it is going to give it back to the host agent. So this is how A2A works. So anyone has any questions, I would be happy to answer. All right. Um so um and then u after uh we have done this uh so all of your questions are in the chat box. So we we will have a 15-minute session for Q&A. So we're going to discuss the rest of your questions uh all of your questions there. So um then how do we uh manage the life cycle of these tasks that these tasks progress through well definfined status that once the task has been submitted it's on working and then the optional input uh which is required if uh so we pass those optional input in it and once the task has been completed or either failed or cancelled the these are the states just like we do polling or just like we uh just like we uh manage the state of a particular API request, we do the same for these task in A2S. So this is called task life cycle management in order in order to keep the track of uh where your task has been gone through to the multiple agents. All right. So moving forward all of the task messages are standardized uh in this particular structure with JSON R RPC 2.0 for request and response payloads ensuring clarity and consistency. It support it support various content type within the messages for multi model interaction whether it is a text files initially or extensible to audio or video. All right. So this client remote agent roles that you have a client agent that actually initiate a task request to another agent and then these remote agent or server agent are responsible for receiving the task processing the task and giving back to the uh client agent. So this has been uh uh so this is the total theory about agent to agent protocol. Let's go and get started with um a a quick uh walk through of a practical exercise on agent to agent interaction. So again if you can see over here we're going to discuss uh there is a uh this is our client which which has a host agent which is going to talk to your uh to two different agent. One of it is trending topics agent and the other is trend analyzer agent. And based on that uh our host is going to actually orchestrate the other agents to provide comprehensive insights. So this is the system that we'll be building to uh today that the trending topic agent analyzer is the one that is going to search the web for current trending topics and the other one is going to analyze uh do a deep analysis with quantitative data and once that is done it is going to give all of these responses back to the host agent and uh your client would give uh the answer to the user. All right. So, um we can actually quickly go over this. So, I have already installed um I have already installed these requirements like uh Google ADK and Uicorn and the rest of these libraries that were needed. So, uh we can skip that. All right. So um so we're we're first and foremost we'll actually configure our environment which is we will be using um a Google ADK library in order to interact build this agent and multiorg orchestration system and these slide decks and uh this notebook would be provided after this and after the end of this session as well right so we're going to import uh from A2A library. We're going to import the client as a rail client module and the card resolver as uh we're going to from the card resolver, we're going to import A2A card resolver which we will be using to create agents card. All right. Then we're going to create a class for patch client module and then we're going to use it to create uh the uh A2A client. So this is more from the syntax how you will actually create the A2A client. After that uh we're going to import async IO for asynchronous operations threading sis OS logging time and all of these method uh all of these um libraries and then we're going to import agent capabilities agent card uh agent skill transport protocol and this is very important when we are creating um a remote agents uh card in order for the host agent to discover that and Then there are uh some other uh um uh some other li uh from some other methods as well that we're going to import from Google ADK and A2A uh repositories. So we import all the necessary requirements. Once that is done, uh we're going to configure our Google API key as well. And then we're going to our our our Gen AI where we're going to make the Vertex AI as false because we'll be using the API key in order to connect with uh Gemini agents. So I'm just checking uh in this uh code cell. I'm just checking from my Google API key what are the models available. I'm going to use Gemini 2.5 flesh uh for the rest of my uh for the rest of my workflows. And then for logging, I'm just uh enabling the basic config with the level name and the message around those lines. And then we'll be building your first A2A system. So we will have to build three agent uh system step by step. So we'll create that first we'll create the trending topics agent that finds the current trending topics and then we're going to uh uh we're going to build an agent uh which is a trend analyzer agent. It analyzes trend with quantitative data orchestrates the other agents sequentially uh which uh host agent is going to do that. So um in in the very first uh in the very first uh uh code cell we're going to create the trending topic agent. So we're going to uh use this variable called trending agent and then we we're going to use uh from ADK Google ADK we're going to use uh the agent method which takes into the account of the model parameter the name of the agent the instruction of that agent uh agent. So here you can see this is the prompt for the instruction of the agent that you are a social media trend analyst. Your job is to search the web for current trending topics. When asked about topics uh and trends extract the three trending topics return them in a JSON format and you must your you must return your response in this particular format. So this is our basically our output parser uh that we are explaining our agent uh our first agent that you should always give your final answer in this way and when asked about a particular topic this is how you're going to search it. So these are uh uh the things that you you're going to do focused on current and realtime trends from the last 24 hours and Gemini 2. Flashflow has a capability to use a Google search. So we in the tools section we are passing Google search. All right. So once this is done um the trending topics agent has been created successfully. Um then we're going to create our second uh we're going to create the agent card for this. And why do we need an agent card? So that uh we can actually give it to the host agent so it can talk to these remote uh agents. So now we with uh we are going to uh use this method from again a library. We are going to pass the variable of name which is a trending topic agent. The URL you can uh you can add any on your local since you're running this on your local host. You can uh give any port to this uh agent. The description of this agent which which means that it searches the web for current trending topics from social media. The version of that agent, the capability uh capabilities of this agent. So we are passing the streaming as two so that uh whenever it gives a response it streams it and the default input mode would be a plain text and the output mode would be a plain text as well but it could extend it to um video images as well as the ADK library supports it. So the uh the preferred transport is JSON R RPC 2.0 that we're going to use for the transport protocol. And then we give the skills that um uh what this agent can do. So this the this is the first agent skill that find trends. The name of this skill is finding trending topics. The description and this is basically a skill what this agent can perform. With this skills the host agent actually identifies uh the host agent actually identify what this agent is capable of. So as we were talking about agent cards, let me go back to the theoretical part where we discussed this how host agent identify and in the first discovery phase. So if you remember we were uh talking about over here that in the discover in the discovery phase and in in the capability assessment um so these two steps are actually uh done with the help of using the agent card. So you created your agent, but your agent card is is the one that actually allows the host agent to identify the capabilities of your remote agent. All right. So with that being said, um moving forward to uh to the next agent uh and then we create a remote agent using the remote A2A agent. uh we give the name of that and the description and the agent card as well. All right. So our first agent has been completed. We did uh what we did we created that agent we created the agent card for it and then we uh deployed it on our local host and created a remote agent. Right? So once that is done uh we can actually uh move towards our second agent which is our trend analyzer agent. So this agent takes a specific trend and perform deep analysis with quantitative data which means that again we're going to use uh the agent method from ADK that uh this is the model name uh model and the model of the LLM model that we're going to use the name of that agent and instructions that we're going to pass to that agent. So you are a data analyst specializing in trend analysis. When giving a trending topic, perform deep research and for each trend you analyze, search for the statistics, numbers and metrics related to the trend and look for in engagement metrics like views, shares, mentions, growth rates and timelines and based on geographic distribution related hash hashtags or keywords provide concrete numbers and data points. Keep it somehow concise. Always prioritize quantitative information over qualitative discussions. And we also give it the tool as Google search. Once that is done, we will actually create an agent card for this agent as well. So similarly following the same instruction that we uh give the name of this agent, the URL where we want to deploy where we want to host our remote agent, the description, the version, the capabilities of that agent and then we give the skill which is the most important part that your host agent is going to identify how this agent is what what are the things this agent is capable of doing. So uh in the in this skill it it is the one that is going to analyze a particular trend. So provides quantitative analysis of a specific trend. So the these are the examples that analyze the climate change trend get metrics for the Taylor Swift trend provide data analysis for AI adoption trend. So these are the examples or the task that this agent can perform. As we can uh as we can see over here that uh the remote agent is the one who is going to create various tasks. So if I if I create a new slide over here and so these are different agent skills. Uh so agent cart is important because your remote agent is the one who is going to create various tasks and based on these skills it can identify. So if we go again and understand the flow um whenever a user asks a query that that query is is sent to the host agent. The host agent takes this query and decompose it, suggest it, plan it and then create multiple tasks for this query. Uh for example, we can take uh we can take an example for this query like uh what is um in 2025 uh what were what were uh the trends um what were the trends related to agentic AI related to agentic AI right so with that being saved so what this host agent would do it will create task one which is um search for agentic AI trends and then the task two would We um like let's say based on a particular agent I tried um uh do deep analysis deep analysis um on maybe a uh agent design patterns AI patterns and then class three would be um analysis on agenti protocols, analysis on agenti protocols etc etc. So all of these could be different tasks that your host agent can define it. So what would be the step one? If anyone can volunteer in the chat maybe you guys can. So step one maybe you guys can uh drop the answer in the chat as well. So um so once the query has been uh dissected into multiple tasks the step one would be our discovery phase in which the host agent would identify the different remote agents available agents. It is going to think about the available uh agents and once the agents have been identified the step number two would be that it is going to uh it is going to do the capability assessment. How it is going to do that? With the help of your agent card and the skills over here, it is going to do the capability assessment. Capability assessment based on the task that is over here versus the skill of that agent versus the skill of this agent. And once that is done uh it is going to uh in step number three uh it is um going to send the task to this uh to the agent and it keeps a track of the life cycle of this task like let's say the task has been uh received by the agent it is currently processing and once that is done it is going to uh send you you're going to get the response back from the agent and you're going to show the analysis to your um to your user. All right. So going back to the slide uh going back to the notebook. So we create a similar analyzer agent card and then we create a remote agent with remote A2A agent method passing the name description and the agent card. All right, once that is done, um we are going to create our agent three which is the host agent or the OS or orchestrator agent which is responsible for orchestrating uh uh uh the number of tasks you have. So the host agent is a sequential agent which means uh the name of this agent is trend analysis host and uh the sub agents this agent has is remote trending agent and remote analyzer agent. So there is your host agent card that it it it is a trend analyst host. The URL of this uh agent is here where you want to deploy it. It orchestrates sequentially trend discovery and analysis using a specialized agent which mean it's a sequential agent uh which means it first it is going to u talk to your host is over here. It is first it is going to talk to the trend uh finder agent which is going to find the trends and once it has responded to it then it will go and talk to the trend analyzer agent which is going to analyze the trend for you. So this is a sequential request. There could be your ho you can make your host agent as parallel as well. you can make them uh in different settings. So once we created this host agent um this the skill of this host agent is a comprehensive trend analysis. Find trending topics and provide deep analysis of the most relevant one. Analyze current trends, what's trending and why is it important? Give me a comprehensive trend report. All right. So once that is done uh we will start uh we will start our A2A servers create a function to run each a agent as an A2A server. So what we want is we want to run all of uh each of this agent as an A2A server. So this is the function which create an A2A server for an ADK agent. We created the agents. we are passing the agent and the agent card so that it can create a server for it. So the it takes into the arguments of these two agent and agent card and returns an A2A starlet application instance which means we'll have a um uh a deployed uh A2A server of this particular agent that will be hosted on our given local host. So um so with the help of this runner method uh we pass the agent name the agent the artifact service the session service and memory service and we config u with the executor config and then we pass them with a in inside a request handler which takes into the account of executor and your in-memory task to store and then it is going to return uh an A2S toilet application instance which takes agent card and HTTP handler which is your request handler. All right. So now we have actually deployed uh uh assume that you have actually deployed your ad AD agents on the given local host. Once that is done, uh we're going to apply nest async io which and this is uh this is where we store server tasks and then we're going to run uh um these asynchronous functions which is run agent server um on this on this particular host and port and then uh we're going to start all of these servers on on the particular given host. So once you run this cell uh your a trending agent would be hosted on this uh local host analyzer agent would be on this local host and host agent would be uh would be actually deployed on this particular uh host. All right, once that is done, all of these servers would run in background and then um you can see over here as I run this notebook, all agents and servers are started and they are running on these particular u on these particular uh local host URLs. All right. So now we're going to test our A2A agents. The two remote agents and the host agents that refers to the uh that refers to those two remote agents are as sub agents. So we will create an A2A simple client that actually talks to the A2A servers. Uh this is your init function that takes into the account of a default timeout. And then um we are going to create a task which is um um this function is going to create a task which is set which is to send a message following the official A2A SDK pattern which means it is going to follow the JSON RPC pattern in order to create task and send them and then we're going to use HTTPX asynchronous client so that we can uh send uh we can talk to uh these servers. is asynchronously and once that is done uh we're going to create um our client and then uh this client would be actually talking to our u A2A servers. All right. So you can see over here our A2A client has been established with the help of A2A simple client and then uh using our um and then you can actually avoid this. This is I was doing to test uh what are the agents or models available with my API key. So maybe I can just remove this. So you can see over here now I've run this uh notebook and uh this is the function test uh trending topics agent uh which actually u so I have actually queried it uh that what's trending today. So um through the Google search it has given me this these answer that these are the trends um Alaska Airlines pilot uh so is Sue Boeing and AI growing influence on social media the description the reason about it AI rapid integration into platforms like Tik Tok Instagram YouTube so all of these are trending topics today so uh this is the answer from our a trend finder agent because it it is being uh hosted on uh 120. And then if we talk to our trend an analyzer agent, it is going to analyze u the trend of AI in social media. So it analyze it. Here's a breakdown of the information I have found and how it maps to the requirements. So it it did all the quantitative analysis of it. I have enough information to construct a comprehensive answer. So it has analyzed a particular trend. Now we are going to test our host agent. So we have asked it find the most relevant trends in the web in the web today. Choose randomly one of the top trends and give me a complete analysis of it with a quantitative data. So basically this agent uh this host agent went ahead it creates multiple task and from there it actually selected the one that Alaska airlines pilot uh s Boeing after midair incident. So this was the trend uh it analyzed it the incident overview the regulatory industry response financial and legal ramifications and all of those things. So you can see over here how uh detailed this notebook was. Uh maybe I can just uh run it with a with a different with a different question as well. Find the most relevant trends in AI uh today and then maybe we can ask it too. All right. So I can run this entire and we'll be sharing this notebook with you guys as well. So you guys can do it uh you guys can run this notebook after our webinar. So all of this is thoroughly tested. So if anyone has any questions I would be happy to answer. Does discovery phase happen after every query or we can cache have uh have these capabilities predetermined or once. Uh Adil that's a really good question I think. Um um how to seek help if we are stuck with collab file? Are there any configuration? There's an API key that we need to be aware of for this file execution. Uh Presa, uh we have I have actually made it more simpler. There are multiple configurations you can connect with Vert.ex AI authenticated but I have actually restricted it with the Google API key so that uh you can just directly go to uh Vert.ex AI studio, Google AI studio and create your API key from here. Uh maybe I can share it uh with you guys as well the link to uh this um all right so let me actually start answering your questions um and let me know if anyone u has any other question related to the notebook or uh the theoretical part. All right. Um, so let me go one by and if you guys uh can re ask your questions if something has been left off. So um all right. So let me start with uh seems like the agent cards have to be published at runtime. Is this true? Meaning a new agent cannot be published after this reason agent reasoning agent is initialized. Um Tim, so basically what um so if you if you have various remote agents that are deployed on uh different URLs, you you can actually add them when you're running your client. So because the discovery phase is started when you ask a query or send a query to your u to your client and after that you can actually go ahead and complete your workflow and then uh you can actually move forward with it. All right. Um so most of the questions are related to notebook and uh uh and the slide deck. Yes, this slide deck and the notebook would be shared and if and most of and most of my students really want the annotated slide decks uh slides as well that the one that I have actually um so this session would be recorded yes you will get the recording too and if you want uh the previous session recording I think because this is a series of webinar we have been discussing agentic protocols with MCPS A2A so the recordings of these of these previous webinars will also be shared with you guys. So you have around I have around uh four or five webinars that are related to each other. Maybe we can share all of them with you that are very correlated. We started with uh love the annotated slide deck. Please share that when as well. Um definitely like this has been a norm that I usually share two versions of slide deck. All right. Um um here it's doing a Google search isn't it? So here how are the results from 2024 or for latest trends? Um I have u I have actually enabled um all right okay I see um um this so it I don't see the notebooks in the discord. Yes k after the session we'll be sharing this notebook as well. Um uh so Aditi for your question um we are for like Google search has a variety of API keys. I am currently using uh the older one. So it is that's why it is not giving the current trends but um so that's why so you can see over here that it is getting the information from 2025 2024. So it's doing the quantitative analysis on the past couple of years. You can say it is not just the data of 2024. Uh Aditi you can see over here it is forecasting it and then so it's based on the model used for getting the results. Yes I have I have actually used Gemini 2.5 flesh. Yes. So it is doing the quantitative analysis of all the previous years and the upcoming years the projection of uh this trend would be so all of these things. Um all right um anyone has any other questions before we so why multiple agents required meaning the why crew agent and agent why not all of the one uh at one type agent? So uh funny this has been u let's say if you like there are some possibilities that you can do directly with nit you don't have to write uh a number of code let's say if you want to create integrations with your Google calendar you can directly do it on uh niten and then you can interact with your host application or the one that you have uh in your native uh so one example of this could be like you have created an lm LLM application in your uh on your company's data and then you want that to be segregated from the one that you create on nitent. So nitent could be your integration layer and then you can use this agent to agent protocol to interact with it. All right into um all right so um any other questions? Uh can we use any version of Gemini in the code? Yes, definitely we can use any version. You can use uh Gemini 2.5 Pro as well. Since it required a a uh it required a a payment method to be configured with Google. Uh since I was actually doing this webinar, I didn't want it to my API key to be like uh after the session, I'll actually revoke it. So that's why I use the free trial method. Um all right. Um oh can the host agent send the same subtask to multiple remote agents and are the subtask done by the human client or the host agent? So subtasks are done by the uh by the host agent. So human human is going to the user is going to give the query and based on that query a host agent is going to divide them into various task and send it to the remote agents and yes it can be sent to let's say two agents are um have the same amount of skills and capabilities. Yes, it can be directed to multiple um sub agents. Can we get this meeting recording link? Yes sar after this session ends you will have this on LinkedIn you will have this on YouTube and if you go to datacind dojo.com uh on the YouTube channel you can find all of my webinars there you will get the slide deck you will get the notebook uh so all of these things would be provided to you um how does it know which task will go to which agent presa that's the question we have been discussing for uh so with the help of agent card of a particular agent since all of these agents have a skills with the help of these skills it analyze which task we should send to which agent. So in the discovery phase based on the agent card we it discovered the number of agents and from the capability assessment it is based on your task versus the agent skill it determines uh what task needs to be sent to which agents. All right so Akmalan how version of uh versioning of these agents are implemented in production how different versions sync for different agents. So uh Akmal think of it like these agents remote agents are now considered as an API. So maybe you update another agent and you don't want your previous agent to be u to be vanished out. So you create different versions of those agents and deploy them on different URLs. All right. Um so for agent orchestration um I think yes A2A has been uh one of the phen it is doing a phenomenal job in like uh orchestrating the task and then performing there. All right. Uh thank you so much Richard for joining. Um and then if you have any other questions you can post it on discord as well. We would be happy to answer. How does the host agent check in code if the other agent is capable of doing a task? Um uh shiv, it is going to do the capability assessment based on the task versus the skills of the agent available. All right. um uh have three stated. I think uh there is a lot of uh questions but um before you guys drop off I would like to hand it over to uh Alishba who is going to actually talk about the details of webinar uh talk about the details of our agent boot camp and I'm going to share my LinkedIn uh profile with you guys so maybe we can connect and if you have any questions we would be happy to answer. So thank you so much everyone and we'll continue our uh our continue our sessions on agenti protocol. Uh we have done MCP A2A and then in the coming sessions we're going to talk about ACP as well. So I'll share my LinkedIn uh account uh my LinkedIn profile with you uh guys so maybe we can connect. All right, this way.