Understanding AI Agents and Agentic Systems

Pretty unforeseen. So, yeah, uh, but today's webinar is understanding AI agents and agentic systems, uh, with Pavan, who's an expert on the topic and is a senior technical evangelist here at Single Store. And we are very excited to have him as a speaker with us today. And, uh, like, and for me, I'm Yupti, and I'll be the host for today's session. One of my main jobs here at Single Store is to bring these webinars, uh, live, like from start to end, to you guys. Uh, obviously, with the help of our amazing speakers. And, uh, so this webinar will obviously have an introduction, and then like, uh, some slides with the subject matter, obviously, which Pavan will be taking up. And after that, there will be like, a, a demo session that I think we have a demo in this webinar as well, right, Pavan?

>> Yes, we do have.

Okay. So, yeah, we'll obviously have a demo in this session, which you can follow along in real time. I will be providing the resources to everyone in the, uh, which who is present in the webinar using the chat button. Uh, just know that you're currently in a listen-only mode, and, uh, we will be communicating, uh, with you using the Zoom Q&A button, which might be present at the bottom of your console. Let me know if you cannot see it. Uh, or you can use the chat button right now. Uh, I'll just enable it for everyone. Uh, so if there's anything, if there's any questions, feedback regarding today's session, feel free to shoot a message to any one of us. Our DMs are always open. And, uh, yeah, towards the end of the session, we will also do a dedicated Q&A, uh, round. So, yeah, like feel free to put in your questions, uh, using the chat button, and we'll obviously take it up live, uh, uh, towards the end. And, uh, yeah, stay tuned for, uh, more information, uh, in the chat button. Single Store is giving $600 worth of free credits if you sign up right now. There's no credit card required, nothing. Just sign up, and you can like, you know, to make the best use of your time, follow along in real time, and, uh, like make the best out of this session. So, yeah, uh, I'll, I'll go dark now, and Pavan, thanks for joining us for our session. So guys, he's an expert on the topic. So if there's anything, do ask the question, and, uh, yeah, keep an eye out at the chat. Thank you so much.

>> Yeah, thanks. Thanks for that intro, and, uh, thanks guys, uh, thanks for joining this webinar. I have a little bit of cold, so, um, excuse me if I sneeze in between. So let's get started. Um, I'll talk about AI agents and agentic applications, and, uh, this is me, Pawan Bagati. I have written a lot of articles on AI, DevOps, and on different, uh, technology topics over the years. And, uh, currently, I am, uh, working as a developer, developer evangelist at Single Store. So that's about me. You can follow me on LinkedIn. I create a lot of, uh, content on LinkedIn, Medium, and, uh, on other websites as well.

So, this is the table of contents for today. We'll talk about the evolution of, uh, AI systems. We'll talk about LLMs and limitations. Uh, the techniques to improve LLMs. Uh, we'll talk about obviously AI agents. Uh, the AI agents architecture. Uh, AI agent categories and core features of AI agents. Uh, agentic AI LangGraph. And at the end, I have a simple, uh, agentic AI demo, a travel agentic AI that basically goes to the, uh, proper websites, uh, scrapes the things, and, uh, comes up with a proper summary, uh, on where you would like to go for a vacation or something like that. I'm going to show that at the end. Uh, but now, let's get started.

I just, uh, went to Google Trends, um, and then saw, uh, the trend for these two keywords, AI agents and agentic. You can also go and search about these two keywords. Uh, you can see right from 2025, the, the growth is just up and above. Everybody wants to build AI agents and agentic applications. Uh, there is a lot of, uh, search about these two keywords, um, globally, right, worldwide. And, um, yeah, you can search this. You can go to Google Trends and try this for any other keyword that you are interested in, maybe MCP or, uh, any other, and it will also show you the geographies, uh, that are interested in and what people are searching about. Um, and again, I went ahead to, um, to this SEMrush, uh, to see the search volume or exact search volume of these two keywords. As you can see, the monthly search volume for AI agents is 27.1K, which is very huge. And, uh, whereas for agentic AI, it is just picking up 49, uh, 5K. Um, sorry, I mean, agentic AI is much more compared to AI agents. Um, and, uh, the keyword difficulty, you can see 98% for AI agents. That means ranking for AI agents is kind of difficult. It's very hard. And whereas for agentic AI, it's also almost the same, like 83%. And also cost per click, you can see $12.59, uh, for AI agents, and it's almost the same for agentic, uh, keyword. So this is about the search volume of both the keywords, just to see how the trend is. People are searching about it. So, yes, um, everybody now wants to build. They want to build AI agents and agentic applications. So let's move further.

This is the evolution of AI systems so far. Uh, if you take, uh, x-axis for capability and y-axis for, uh, autonomy, it all started with LLMs, basic LLMs, with no custom, uh, knowledge base attached. You know, uh, LLMs were trained on a lot of, uh, raw data, structured, unstructured data, anything on the internet. Uh, they did not have any context. Uh, you gave some input query, and then they just produced the output. And then came RAG, where you augmented your own custom data, uh, so that LLMs can't, you know, uh, can easily go fetch, uh, your own data and come up with a proper, contextually relevant responses back, right? And then came AI agents, where the autonomy principle came in, right, um, where, uh, you had these agents going out and doing their own stuff, uh, uh, whenever a user query comes in, uh, they can take their own decisions, which, which tool to call, where to go, should I do the API calling, tool calling, like how to fetch, fetch the answers, and all those things. And then came agentic systems, where a lot of AI agents came together to build, to help you, um, you know, build these agentic, uh, systems. We'll talk about, uh, agentic systems and agents more, but I just wanted to show you how, where we are right now, and it is, this, this evolution is still moving up. And, uh, but let's start with the basics, obviously, with LLMs. Uh, what are LLMs? Basically, they are made up of structured, unstructured data, and the architecture, the Transformer architecture, basically proposed by Google folks, the researchers at Google, and then some kind of training, right, um, um, that's how, uh, the LLMs are basically made up of. But, uh, these LLMs, if you use them just right, right away for your AI applications, they don't yield good results because they have limitations. They have knowledge cut-off dates. We all know that every LLM has a knowledge cut-off date. They are trained on till a particular date. If you ask anything that's happened recently, now they may, because now they can easily do tool calling and all, but they usually, uh, have this knowledge cut-off, and that's why they basically hallucinate. They come up with their own answers and sometimes made-up answers because they don't, they can't verify their own answers. And, uh, they don't have, they don't have the, um, um, you know, a connection to your, uh, custom data, uh, right? Only once you give them that data, they can have the context. And there's also context limitations. They, they can only, uh, handle, um, you know, to a certain point, right? They, they don't, uh, they start to, um, uh, what to say, um, don't understand some of the things, and they don't have, uh, they have this context window size to remember only till that point. Uh, apart from that, I know after that, if you feed in a lot of data, they start, uh, losing the context of what you're saying.

So, these are some of the limitations, and these are the techniques to overcome these, to improve LLMs, the basically the hallucinations of LLMs. Um, one is prompt engineering, which is nothing but the art and science of talking to your LLMs. The, it all starts with the user query, like, um, there are a few short prompting, CoT, chain of thought, you might have seen in LangChain, right? So, it's, it's all about how you, uh, use your query and talk, the art of talking to the LLM. And then comes the fine-tuning, which is basically you take a base model and, uh, you train that with your, uh, you train the base model with your, uh, custom data, and that's how you get a fine-tuned, uh, uh, model that you can use for your, um, uh, applications. But, uh, fine-tuning costs a lot of, uh, money because there's a lot of, uh, computing resources required. But actually, um, when it comes to improving LLMs, RAG is the recommended one. Uh, retrieval augmented generation is what we call, because here you use an external database, we, what we call usually the vector database, and you augment your custom data through that database, so that the LLM can easily go and fetch the top, uh, chunks, and get you the contextually relevant responses. So, RAG is the recommended one, um, in improving the LLMs. But they all share a common limitation. What is that? They only optimize a single shot, single shot of reasoning. What does that mean? We usually go to ChatGPT, you might have seen, like we ask one prompt, they, it doesn't think by itself and say, "Hey, I think I did this wrong. Let me go back again and correct this." Right? We always tell ChatGPT, like, uh, to do this, change this, update this, update that. So it takes a lot of time. And that is where the transition happened from LLM to AI agents, where agents are systems that independently accomplish tasks on your behalf. So they can easily pursue what the user is asking. They can reason, they can act by themselves. They, they, they will get to know what the user is asking. So they can take, they can easily take the actions, do tool calling, and all those things. So that is where the, um, autonomy principle comes in for AI agents.

So, this is a simple AI agents architecture. Um, there are LLMs, there are tools attached to it, and it can do planning, right? The, do goal decomposition. If the goal is, if the, uh, user is asking a very complex problem, it can basically decompose that whole task into subtasks, and it can also do reflection and critique its own responses, and, uh, yeah, there is a memory also, short-term and long-term memory. So, this, this is, these are some of the simple components of an AI agent. And here is an AI agent in a code format of a weather agent, you can call it. The name is weather agent, and there's an instruction, right, prompt instruction, "You are a helpful agent who can talk to users about the weather." And, uh, yeah, you, whenever the user asks about, "What is the weather in San Francisco?" and this particular, the agent can trigger this tool calling, uh, and get the weather from San Francisco.

So, uh, there are categories. These are the AI agent categories: the single agent system and multi-agent system. As the name suggests, there is, in a single agent system, there is one agent connected to a lot of tools. And whereas in the multi-agent system, there are multiple AI agents coming together to solve a complex problem, right? So here, only one AI agent in a single agent system does everything. And it is okay when, uh, to use a single agent system when the user query is very simple. And, uh, and you, you can go for a multi-agent system when the user query is very complex, and you are solving a very complex problem.

So, these are core characteristics of an AI agent, basically the autonomy, right, uh, minimal, uh, human intervention. That's why AI agents are all about, like, we talk about the agency and autonomy. Task specificity and narrow, well-defined tasks. And, uh, the reactivity, responding to changes. It's not linear, it's very dynamic because the user can come up with different questions, and the AI agent, it, and should easily, should be, should easily sense what he's asking, and it can, it should do tool calling depending on the user query. So, there's a lot of changes involved. So, there's a lot of, uh, dynamics. So, it's not linear. So, these are the three basic characteristics of the agents.

Here is a simple example of an AI tutorial assistant. Uh, the, if you take the whole thing as an AI agent, uh, you can see there is an LLM attached, there's a system message, there's an orchestration layer where we're using LangChain. LangChain is again a framework, uh, to build LLM-powered applications. And, uh, there are tools, there is a knowledge base, and there is a memory, right, there. And there is a UI. UI is where the student can basically come and ask, uh, different questions. And there is an LLM, for example, as your OpenAI GPT guru. And, uh, the system message is, "You are an AI student tutor, and your role is that of helping students with blah blah blah," and, uh, you know, uh, understanding complex STEM problems, topics. And, uh, so there is basically LangChain handling everything, connecting everything together. And, uh, in the memory, you can see the student's past exams, their scores, their strengths, and weaknesses. And, uh, in a knowledge base, you can see school's reference manuals, articles, and, uh, all the course books, everything. And, uh, uh, the student can easily go pick, pick the school's calendar in the tools and then book his next, uh, uh, tutor to class or, uh, exam. So, a simple example, right? Um, and, uh, this is a simple, um, uh, differentiation between LLMs versus AI agents. LLMs, basically, they predict the next word, right? They are only able to predict the next word, whereas AI agents do a lot of things. They, they have these tools, they, they have this memory, they have, you know, they can take actions, um, they can do the planning, they can easily sense the environment, or all those things.

So, this is the journey so far to agents. We had only a single LLM, uh, doing everything. Then came workflows, where we, uh, try to talk about RAG, retrieval augmented generation. And then came agents, that took their own, you know, they started taking their own decisions and taking their own actions, uh, through planning and tool calling and all those things. And again, the same thing, from LLM to RAG to agent. This is a simple journey. And, uh, yeah, uh, the future is, the future of AI is agentic. Let's talk about agentic AI.

So, when we talk about agentic AI, um, you can see it's important to understand the word, to understand the meaning of the word "agentic." It means the ability to act independently and take decisions, right? So, and, uh, when it comes to agentic companies on the growth trajectory, you can see Lovable. I think most of us might have already used Lovable. It is one of the fastest-growing companies, agentic companies, which grew from, you know, 1 million to 100 million ARR in just, I think, 10 months. So, after that, you can see Cursor, with Open, Slack, Will, and Shopify, coming up. And, uh, yeah, this is, this is some of the funding news from 2024 to 2025, right? Agent startup Emergent raises $23 million, agent startup Adopt AI raises $6 million, $1.2 million, $7 million, $25 million, and some of the news from different publications.

So, this is the evolution of agentic AI. We started from AI systems, using simple, you know, uh, simple LLMs with simple rule-based tasks, low autonomy, low agency. And then we started building simple AI agents, a little bit of specialization. And then task-specific autonomous, uh, agents. And then came collaborative AI agents, where multiple AI agents came together. And then now, AI agent ecosystem, where collaboration between multiple applications and organizations are happening. That's what we are seeing. And this is from Gartner. Uh, and, uh, the goal of agentic is about increasing agency and specialization, right? As you can see, if you take x-axis for specialization and y-axis for agency, you can go to ChatGPT and say, "Create a meal plan for me," and it'll create, which has less specialization and less agency. And, uh, you can also create a personal assistant. Uh, you can have a personal assistant and tell him or her, "Here is, uh, no, create a meal plan for me," and it'll come and say, "Here is a meal plan for you." It has more agency, but, uh, less specialization. Whereas, when you have, uh, you take the example of a personal grad student who is specialized in nutrition, uh, he can say, like, "Evaluate everything and say, the nutrition of, uh, your meal is this, and this is what you should eat, and this is the meal plan for you." And he or she will have more specialization, plus, but less agency. He or she can do only that much. Uh, you can't do much more. But when you have a, a agentic team, you can do a lot more, right? As you can see on the upper right corner, "Here is your meal plan. It is, uh, dairy-free since, uh, blah blah blah is visiting, and, uh, groceries will be delivered tomorrow at 8 p.m." You can do much more with, if you're having an agentic team, that means having different AI agents, uh, for this particular thing, that, and it will have, it will have a lot of more agency and more specialization.

And this is the journey to agentic systems, from again, LLMs with stateless, long-running, computationally intensive resources, to agents, the structured, uh, enrichment loop that builds context, moves tools, takes actions, and, uh, does a lot of things. And then now we see agentic systems, where networks of multiple agents orchestrated to solve complex problems and tasks, right? And, uh, this is a simple, um, example, I mean, differentiation between AI agents and agentic AI. AI agents, as you can see, there's one agent which is connected to a lot of the tools, the data, different tools. Whereas agentic is different, a lot of AI agents coming together to solve a complex problem. So, basically, in AI agents, there is one AI agent that acts as an orchestrator. Whereas in agentic AI, there are different multiple AI agents come together to solve a complex problem. You, we can see this in this, uh, use case, right? So, if you take an agent, ask it like, "Hey, I want to book a trip, uh, to, um, wherever." Right? So it'll say, "Sure, I will find the flights for you." It'll just do and come up with the flights, that's it. But with agentic AI, you can do much more. "I want to book a trip." So, one agent will go find the flights, one agent will go find the hotels, and one agent will basically come with the summary plan for, and which is in accordance with your calendar, right? So, that way, agentic is much more superior, right?

And, uh, this is the tool usage in agentic. The user prompt comes in, and he'll ask like, "Hey, I want to know what is the weather in Mumbai." So, um, and then the LLM basically picks the right tool, uh, from the available toolset. It'll pick the weather tool, and the tool execution happens, and again, it goes to the LLM to give the response in the natural language, um, and that finally, the output will be, um, uh, produced. And, uh, here's a simple RAG workflow, uh, because I want to give an example of an agentic application, that is, I want to show agentic RAG application. So, here is a simple RAG workflow where user query comes in. Um, sorry, let's talk with, uh, our own custom document, that is, PDF gets converted into embeddings using an embedding model, and then every, all these embeddings are stored in a vector DB. You, you see the down line here, right? So, all the vector embeddings, all the embeddings are stored in a vector DB like Single Store. And then when a user query comes in, even the user query gets converted into, uh, embedding using an embedding model, and it goes to the vector DB, uh, using vector search, uh, methodology, and then it'll fetch the relevant context. Now, you have the query and the context. You, it will go through at the end to, uh, an LLM to finally, uh, respond back, right, to the user. So, there is an absence of AI agent here. But when we talk about agentic RAG, here you can see the agent coming in between to take, which to, to see, you know, which tool to call. Above everything is RAG, and, uh, there is one addition here, that is the agent that will basically do everything here, right, which tool to call. And, uh, apart from that, everything is same. So, that is where agentic AI, agentic RAG is, uh, becoming popular these days. And, uh, this is a multi-AI agentic system where, uh, multiple agents coming together to solve a complex problem, like, find a flight to New York, right? The, there is an orchestrator agent that breaks the, um, user query into subtasks, and, uh, each task is, go, each task goes to, uh, different, uh, specialized, uh, agents. The planning agent will determine the best flight option. There's a booking agent, memory agent, there is a critic agent to see if we have, we are doing everything right. And then there is again, at the end, there is a, uh, agent that, that basically puts out the summary to the user, saying, "Hey, this is your final output, and this is the report. This is the final, you know, these are the things that you should do. That these are the flights, best flights we found for you to go to New York." So, this is the difference. A simple differentiation between generative AI and agentic AI. Generative AI is where, you know, it understands and generates content, that's it. Whereas agentic AI, it understands, generates, and performs actions, because it can do tool calling and it can take its own actions. And, uh, yeah, traditional generative AI, where, um, you know, there is an LLM, basically predicts the next word, that's it. Whereas agentic AI, uh, there is an agent in between. The goal is, "Build a web app using React JS," and then it can reason, it can, it has memory, it has tools, it can do research, and then come up with a proper code and tell you, uh, yeah, this is how you can build this yourself, right? So, that is the differentiation between traditional generative AI and agent AI.

So, there are different agentic frameworks. In this tutorial today, I'm using LangGraph. There are many others, but, uh, yeah, LangGraph is a graph-based workflows, real-time processing, structured data handling, and, uh, it is best fit for agent type, like, you know, reflex-based, model-based, utility-based. Whereas CrewAI, you can use CrewAI where, uh, there's a role-based design, task delegation, and multi-agent collaboration is required. And AutoGen is from Microsoft. Uh, it's conversational approach, quick responses, modular design, and supports planning. And, uh, basically, now, since, uh, LangChain has the first more advantage, most of the people are using LangGraph. And then CrewAI. I really like LangGraph. So, again, these are some of the other, uh, frameworks that you can use, uh, with code. And there is, no-code or low-code platforms. With code, LangGraph, uh, Agno, Agents SDK from OpenAI, Agentic AI2, right? AI Google SDK. And then no-code, low-code, you can see Botpress, Make, Anything is making a lot of noise these days, right? Everybody is building. It's just a pick and drop, uh, thing. And LangFlow, ChatFlow, WatsonX. And these are some of the low-code, no-code, uh, agentic frameworks. Since I'm using LangGraph, so LangGraph is passed, is, is a part of LangChain again, LangChain ecosystem. So, LangChain is basically used when there is a predictable output, like A, B, C, right? Very linear, uh, workflow. Um, and then LangGraph is for dynamic workflows, right? Uh, that is where, uh, we build AI and agentic applications using LangGraph, where, you know, the, once the user, it can go from A to B, or, uh, if it thinks that, "Hey, before going to B, I need to go to C," it can do that also in LangGraph. But that can't happen in LangChain. It has to be from A to B to C, very cyclic. So, that is a simple difference. U, and then these are some of the components you need to know: nodes, edges, and state. Nodes is like what the agent does, like LLM calls, tools, logic, and all those things. And edges is like how it decides what to do next. And state is what it remembers and carries forward. As simple as that, right? And, uh, so the, so the journey so far is like, uh, the LLM workflow to, we went to RAG, uh, to handle our custom applications with our custom data, and then we, uh, saw, uh, you know, autonomous AI agents coming, and then now we have agentic AI, where different AI agents come together to solve a complex problem, right? But, uh, yeah, this is a, again, a typical enterprise data architecture. Uh, you, you, you see this in most of the enterprises, where they will be using a lot of databases, right, which is a real mess. First, they'll be, you know, they think that their customers are running a lot of analytical workloads, they will have an OLAP database. And after some time, they, and and some customer will come and say, "Hey, they have a lot of transactional workloads, and they feel like having an OLTP database." ASE. And now, I know, cache, for cache also, one database. And then at a certain point, you know, now the whole world is talking about AI applications, building AI applications, and now they will also have a vector database, which is a kind of a mess for to handle, you know, for your architect to, to handle all these databases at once. So, at Single Store, what we propose is like, you know, you can have Single Store as an all-in-one data platform that can ingest and index documents in, uh, you know, milliseconds, and can do re-ranking, similarity matching, you know, uh, you don't have to have a specialized vector database just to have store your vector embeddings, because Single Store supports all types of data. So, so that's, that's the thing what we propose. And Single Store, uh, helps you build robust, uh, yeah, applications, because we have, uh, amazing integration with all these, uh, popular frameworks, LangChain, LlamaIndex, CrewAI, and free data. And, uh, yeah, so this is the agentic AI travel app that I'm going to show you today. So, the user input is like, um, uh, I want to go somewhere, and these are the dates, and my budget is this. The bedroom, I want one bedroom, two bedroom. And, uh, you have the accommodation agent, and we have the flight agent. We have two agents that will go and scrape the particular proper websites. Uh, here I'm using dummy websites in my, uh, tutorial. So, at the end, uh, the everything will be summarized and given a final report to the user, saying, "Hey, this is, this is your final, uh, vacation report. You, you can go here, and this is the bedroom, this is the hotel you can choose, and this is, these are the flights you need to book." So, let's see that.

I'm using LangGraph here to build all this. So, this is a travel planner. Think of this as a travel planner. And this is my code structure. We have different agents, right? We're using basically at the top, LangGraph, framework. We have different agents, cache agent, agent. We have flight agent, recommendation agent, search agent. We have the graph.py where it defines the flow of the entire agentic app, right? And then the state.py where central memory box of the agentic system. And the database holds all the database logic. That is, we are using Single Store here. And, uh, yeah, we have the main.py. It basically runs the entire application, everything. And, uh, so let's go to. So, for this, you need to have a Single Store account. So, get a free Single Store account. Single Store is a free tier. And you can sign up, sign in, and you need to create a workspace and a database. And then you need to create two tables, that is accommodation and flights. And then you need to start the application. Let me show you that.

So, this is the portal. Uh, once you go to Single Store, sign up and sign in, this is where you will land, basically, in the deployments. And if you don't have a workspace, create one. And then attached to that, create a database. And this is my database, travel_agreed. I, I already have a workspace. And after that, you need to create two databases, sorry, two tables inside a database to hold flights, and then accommodations, right? And you will go here to Editor, and go here, and then New SQL file. This is where you will create those tables inside your database. So, how you will do that? You need to select your, uh, connections. So, this is my workspace, and I'll also select my database, travel_ag, and here. And, uh, let me, uh, go to this, and let's go to the. So, this is my repo. I have already, uh, shared it on my GitHub, and everything is here. You can follow this. I'll share this. And this is, you will copy this, this particular SQL query, and then that is for accommodations table, and this is for flights table, right? And we'll copy both, we'll paste it here, and we'll just say run, and those two tables will be created inside our database. And, uh, I can also show you how it is created. Travel_ag. So, both are created already. Think that both are created. You can follow this along and do that. And then the next thing is what we need to do. We need to install the, uh, requirements. And you need to share this, create a .env file and add all this, uh, to host, user, password, DB. You can get it easily from here. If you go to here, and then go to SQL ID, you will get all these things from here. Okay. And after that, we'll have created this. We'll run the app. So, we'll go here. I've already added it. Python app, python main.py. If I run, ask like, "Okay, where I want to go, uh, from origin BLR. Then I want to go to Mumbai. And then the date is 206. Month is 02. Date is like maybe 14. And then again, 2026, 02. I would like to come back, and it can be, uh, 18. Bedrooms, one. Um, max price per night, I'll say $100. Yeah. Yeah."

>> And then minimum rating four.

Now it is generating. It will generate and it will give us the, it should give us the output. So, yeah, it's taking some time. Yeah. Right. It, it gave us the, the accommodations. Uh, and then the, and then the weather. Temperature range is between 24.2 degrees Celsius and 28, 28.4. And the top hotels for Mumbai. And these are the top flights, right? And, uh, yeah, some of them are little dummy. I did not use the proper, uh, uh, websites, but you get the thing, right? You can convert this properly to scrape the websites. I did not scrape those websites. So, yeah, these are the agents, like I said. So, that's how you can basically, uh, you know, we have the weather, uh, agent, we have the, um, all these agents, and we also have a cache agent. If you ask the same thing again, the output will be very fast, right? So, that's how you can basically build. I'm going to share this, uh, in the chat. Uh, also. Okay, I think you will share that. I'm not able to share that. Okay. It's here. Check. I'll share this so that you can share. Yeah. And, uh, yeah, I think we have come to the. Sorry, that's it. That's how you can basically build agentic applications. Let me know if you have any questions. You can share that repo, you think today, these people.

>> Right. Sounds good. So, uh, Pavan, we do have some questions for you that are in the Q&A.

>> Q&A. Q&A.

So, folks, for everyone that is here with us live, we are officially starting the Q&A session. And, uh, you can use the Zoom Q&A button, which is present at the bottom of your console. And, uh, you can put in your questions, and Pavan is here to take them up live. So, uh, our first question is from Harish Das. They say that, "From what I understand..."

>> Oh, yeah. Okay. Amazing. So, if you,

>> prefer that, that, uh, that's good too. Okay.

>> Yeah. "From what I understand, the LLM is the brain of agents, yes, uh, and agentic systems. If so, how dependent are AI agents on the LLMs they use, and how easy is it to switch LLMs and AI agents use if needed?"

Yeah, you can always switch, uh, your LLMs. But, yeah, that is true. LLMs are considered as the brain of these agentic or AI agents. Uh, so how dependent are AI agents on the LLMs? They are the brain. So, you need an LLM there. If you build any AI agent, you might have seen in the code, either OpenAI or any other LLM people use. So, you need an LLM, and that is also the brain that decides which tool to pick. Uh, so that also comes in the, you know, the reasoning thing in the AI agent, right? So, it, it acts as the reasoning layer, uh, because that is the brain, like you rightly said. Yes. And you can also switch it. Like, you can mention, uh, which LLM you want to use, and you can change it if you want, as many times as you want.

>> Do you have any recommendation for books, online courses for learning agent systems?

So, there is this, uh, DeepLearning.AI. Yeah, I think, um, from Andrew, uh, NG, I think it's the best. You can go there. And also, uh, uh, there is some one YouTube channel, I don't remember now. AI Engineer. AI Engineer YouTube channel is really good. They talk a lot about, uh, AI agents and agent applications and all the MCP, all the trending, whatever is trending in AI, you can find in AI, uh, engineering. Yeah, I think the YouTube channel is AI Engineer or something like that. Yes. Let me stop sharing so that I can go and search if that is the one that I'm right. If I'm, AI. Yeah, it is AI Engineer. If you go to AI Engineer, you will see all, uh, meaning amazing videos by real folks, actually. Okay.

>> So, how do you write and run unit tests for the agent codebase? Can we spin up and run the agents locally for testing?

I mean, for locally, there is, uh, this thing, no, what do you call? But, uh, do you write and run unit test? Unit test, now everything is automated, right? Because, now we have, what is this, CodeRabbit, and all. Uh, they can easily, there are different AI tools that can easily, uh, write these unit tests for you. So, I don't know, now, I, I don't know after, might be very controversial, after AI, I don't know about the, the testing folks and all. Um, so, yeah, you can use AI tools to run your unit, you know, uh, unit tests. And there are tools like CodeRabbit and all, which can review your, uh, code easily. So, yeah, there are tools like that which you can use, uh, locally for testing. Yeah, you can also try that. I have not tried. Can we spin up the and run the agents locally for testing? I've not done that. Maybe using Ollama, you can run, do something, connect your, uh, GitHub, that repo, particular repo, and run that locally. Yeah, maybe.

>> Does your demo app have guardrails?

No, I did not use guardrails here. But I have a video, I have made on my YouTube channel where I have shown like how guardrails should be used to make sure your LLM doesn't hallucinate and behaves as expected. This is just a demo. I just wanted to show you guys how different AI agents come together, uh, and give you a proper output, right? So, I did not even scrape proper websites there. It was just a, because scraping sometimes is not permitted. Uh, right? I don't want to show you to me scraping a website, uh, proper website. So, if not, are there any frameworks that help with that? Demo guardrails, uh, if you're talking about guardrails, the Nemo, sorry, Nemo, Nemo from Nvidia, you can use those. And there are many other, with, with LLM only, you can create guards, like LLM as a judge, a proper LLM as a judge, which can, you know, actually critique its own, uh, response. That way. So, yeah.

>> Can you implement knowledge graphs with Single Store for building structured memory for a?

Amazing, amazing. So, yes, we can. Since we have SQL support, you can build knowledge graphs using Single Store. So, I recently have created, made a video where you can use Single Store for your traditional RAG, as well as your, uh, Graph RAG. So, in Graph RAG, what you do, you, you know, you need to have, you need to create knowledge graphs, that's what you inject, just for proper entity relationships. And, uh, yeah, by the way, Graph RAG is mostly used these days because of its, uh, efficiency. But, yeah, for your question, Single Store can be used for both. If you go to Single Store YouTube channel and type Graph RAG, one of our engineers has made a, uh, video webinar on how to, how you can use Single Store for, uh, for your Graph RAG to create Graph RAG applications. And I have also made a video. So, yep, Single Store Graph. If you search together these two keywords on Google, I think you will get it. And we also have a, I have a blog written on.

>> Thanks. Thanks, Harish. Thanks. Thanks, guys. If you have anything, I'm very active on LinkedIn. So, you can go check out and, uh, yeah, I talk a lot about AI there.

>> All right. So, thanks, Pavan, for doing this amazing session. It was, uh, iterative. And, uh, for other folks that are still there with us, first of all, thanks for joining us today. And, uh, we have amazing sessions that are coming up. So, uh, Harish, we will share all the, uh, I just see your question. He's basically asking, "Will you share the slides?" The recording, the slides, all the resources that are relevant to the session will be shared later in a follow-up email. So, stay tuned for that. And we have an amazing webinar tomorrow as well. So, that's a different topic. It's on governance. And I will also mention that in our follow-up email as well. So, basically, it's about, uh, identifying risky audit logs. Uh, so if you're interested in any of any, in any such topics, you're more than welcome to, you know, uh, join us there, because it's a very special session, and a lot of time and effort has gone into preparing the content for that. So, uh, yeah, stay tuned for that. And tomorrow is a, a good session that we are bringing. And there's more sessions in the future as well. So, we highly encourage you all to keep an eye out at our, uh, events website. So, there's a section for webinars. You can go scroll through that. And I'll make sure to add it in our follow-up emails as well, if anyone's interested. So, yeah, uh, I hope you enjoyed today's session. And again, if there's any questions, you can always DM us, and we'll be very happy to get any feedback or any suggestions for future topics. You're more than welcome. You're more than welcome to let us know. And, uh, yeah, I hope you all enjoy today's session. And thanks again, Pavan, for putting it together. And, uh, I hope everyone has a great rest of the.