What is an AI agent

0:00 Let us begin with some theoretical foundation. This is their website. They market it as multi- aent platform. Crew means you have multiple agents working together. They have a way to create crew and agents through some UI. They call it UI studio. We are mainly going to be focusing on the framework which is by writing code. We will be creating the crews. They have around 30,000 stars on GitHub which is pretty good. So now since it is a multi- aent platform first we need to understand what is agent.

0:47 Uh so let's spend few minutes understanding what is agent and agentic AI and if you already know this fundamentals then you can skip this section and move to coding part. So what exactly is agent and agentic AI when you're using chat GPT behind the scene you have LLM as your main component. This LLM can be GPT claude Gemini anything. when you ask a question, it will generate answer based on the pre-trained knowledge that it has. That is why it's called generative AI.

1:10 But let's say if you ask a question, what is the price of Emirates flight from place A to B tomorrow? It won't be able to answer because it has a knowledge cutff. Okay. So this is generative AI with only LLM GPT4 claude whatever model you are using it will have knowledge cutff that it's trained until that point of time and any event that happen afterwards it won't have the knowledge but you can give it an access to some external API let's say if you give uh it an access to Xedia API or make my trip API okay some travel API then it will be able to call that API and pull the latest price.

1:57 Okay, you can have a different question such as book me cheapest flight from New York to Delhi. And now if LLM has access to Xedia, it will call Expedia for that you know source A to source destination, source to destination and given date. It will then search for the flights. It will decide which is the cheapest one and then it will also book it. So here it's not like chat GPT like simple question and answer. It is actually taking action for you.

2:22 Okay. And it will say okay flight booked. So this is called AI agent. Okay. AI agent is a program that takes inputs, thinks about it. It acts to complete a task using tools, memory and knowledge. The API that we are using Expedia is called tool. Okay. And it is autonomous. Autonomous means it is taking action for you but narrow task. Okay. So here see it's just simple flight booking. So it's a very narrow task. So if you think about it, it first searched for few flights.

3:11 Let's say it found five flights. Then it decided which is the cheapest one. Let's say this is the cheapest one. So it decided that in autonomous fashion. Okay, it did not ask you. It decided that hey this is the cheapest flight and it also booked it. You see? So that is why we are saying it is autonomous. Okay. So that is your AI agent. Now you can expand this further and you can say book me a flight for my 7-day trip to New Delhi in May and I personally like sunny weather.

3:35 So I also have a requirement that the weather should be sunny all the days and flight budget is this no layover. Now LLM can take help of Xedia API for the flight information but it also wants to know the weather. So maybe for weather it calls acue weather or weather.com whatever API right Google weather etc and it will pull that information. So this llm can have access to multiple tools. Okay. So it is checking the weather first to find seven consecutive days in May where the weather is sunny and for those days it is comparing the airlines and finding a flight which is in your budget.

3:59 And not only that, it can also go one step further and give you suggestions for hotel and airport taxi. If you have booked a flight from make my trip or Expedia, you will find this, right? They will also give suggestions. So then it will say okay flight booked from this date to this date, weather is sunny, I also have good suggestion. Okay, so this is a flight booking AI agent. You can also have a system with multiple agents.

4:48 For example, you have flight booking agent but if you are not let's say if you're traveling to India, if you're not Indian citizen, you have to take visa, right? So you can have immigration AI agent which you can connect to your one drive. You can have your passport, your immigration document there and this LLM can access it and it can tell you that hey you don't need have a visa. Okay, forget about uh booking the flight ticket.

5:11 You first need to get the visa. So the this system consists of multiple agents. It is doing multi-step reasoning. It has a multi-turn approach where it is calling multiple agents. Maybe agents are being called multiple times too. And it is multi-step planning. It is working on a complex goal autonomously. Okay, you just giving it a goal that okay book a flight ticket. But the system is checking your visa. It's like if you call a agent, right, travel agent, they will do all of this.

5:35 So this is similar system and this is called agentic AI. Agentic AI is a system where one or more see one or more AI agents work autonomously often over long task. They might call LLM multiple times making decision using tools and calling even other agents to reach a goal and crew AI helps you build AI agents especially multi- aent system. Now there are a lot of other frameworks right Google ADK lang graph crew AI so what is the difference when should I use what if you're building lightweight fast agent let's say if you want a agent which performs really fast and if you want to build it quickly agno is the framework Google you will use if you want tight GCP integration okay let's say your company is in Google ecosystem and and you want to leverage Gemini and all that other ecosystem Google is good if you want graph-based stateful orchestration which is highly customizable, highly reliable agents

6:46 then you use langraph and crewi is used for role based orchestration simple and less customization. Okay, as we write code you will know more you know the the power of all these different frameworks we have separate tutorials for agno lang graph etc. So you can try it out and one of the frameworks can work the best for your given use case. We are going to use