What is an AI agent?

One question that keeps coming to me, what are AI agents? The definition for AI agents looks pretty flexible these days. It depends on whether it's just a bot, it can be just a chat agent, or it can be like anything. That's why I created this first flow, that when you're talking about how one agent works, we will be able to standardize that definition with this kind of a reasoning loop. Whenever you're building any kind of an agentic system, either it can be a one agent, it can be a multi-agent. You will be seeing these kind of five or six key stages as a part of it when you're building your thought process. The first thing, it starts with a user goal. What kind of questions, what kind of a task that you want that particular agent to target and to work upon. The second step it's going to be, you will be having some sort of an LLM. You can call it as this is going to be the brain of that particular system, that plan steps, that breaks down risk, that defines that, hey, how that user goal can be achieved and based on it, it will start taking those kind of an actions. To take those actions, definitely just like as a human, we need some tools. Agents will also need some sort of a tools. If you're searching, you might be requiring Google search or a Bing search, or if you have a database. For an example, I read a lot. I have a lot of papers with me. Whenever I read a paper, I download it and upload it back to my Google Drive. That acts as a repository for me. So that can be like another tool that I need to verify, that I need to cater to, that I need to parse when I'm building my understanding. The next step is going to be the agent observes when it's parsing any details, when it's learning from those kinds of tools, it's getting those information, those tools, that what kind of results it generates and based on that, it self-corrects. Am I going in the right direction or not? Or is it what user asked for? Is it aligned to that particular user goal? And based on it, it's keep iterating. Sometimes we add this iteration as a dedicated step. Sometimes it gets added as a part of the observations of the observed step itself. But I kept it separately because this is really important. Till the time loop completes, you have your final answers. And the final stage is the consumption or you can call it as a delivery. Is it going to be in form of like just in the chat? Either it can be an e-mail so that it can run automatically in the backend. It can be a podcast. It can be a video clip. It can be anything depending upon what the user goal was. This is what we call it as the reasoning loop. So whenever we are going to build any kind of an agent, we need to go through this kind of a loop, starting from user goal to go for the final delivery. The quick question comes in, why agent is getting so popular? Are there any numbers that it can quantify where the industry is evolving? So the three key things I just want to share it with you. The first thing is about the market potential. Based on the latest report by SCADISTA, it's going to be $47 billion by 2030. It's exploding with a rate of 45% CAGR. That is a compounded annual growth rate. That's a pretty, pretty big number in one technology itself. The second one it's about, that's published by Gartner, and they said that almost one third of the enterprise software will be including an agent in KI. Either they are going to be direct agents, or they are going to use some flavor of the agents as a part of the software. And right now for the 2026, if we think it through, it's acting as a tipping point. Either it can be hyperscalers like a Google or a Microsoft, or some of the largest model developers, OpenAI, Anthropic. Everybody's targeting launching certain kind of an agentic frameworks. It can be agentic platform. Think about like the latest Anthropic Core work. Think about OpenAI Frontier. Everybody's talking about building end-to-end platform, definitely more on the enterprise side, but that's going to be very similar on the consumer side. How can you build an agent? How can you integrate an agent? How can you get different tool access? How can you evaluate and observe your agent? How can you consume those agents on a real time basis, on a daily basis? If I think about it, the question comes in, how can you build those agents? If I'm you, I don't have any kind of a technical skills. Is it really tough to build those agents? So I categorize building those agents on three dimensions, and it's surprisingly very easy to build an agent. The first thing is a no code. you do not need to write a single line of code. Just write some prompts, instructions, use step-by-step guides. In minutes, you can deploy that agent and start consuming it. Some examples, Azure copilot Studio. No engineering needed. Just go for it, define your goal. They will be having like a one chat bot itself where you can keep defining your goal and it will keep creating agent for you. Google Agent Builders, like a pretty new offering from Google itself, where you can create your workflows, drag and drop, build those things out, very simple and easy to push for it. AWS Party Rock. These are some of the tools you can find from the hyperscalers, but you will be getting tools across different products like Loveable, talking about Figma, talking about Adobe. There are multiple tools available that you can create your agents on, depending upon what kind of a task that you want to achieve. Second one is a low code. I don't know how to code, but I can write little bit of code. So you can go for Vortex CI agents on Google, You can go for Bedrock Agents in AWS. You can explore OpenAI Assistants, or you can actually explore AI Foundry. That's a newer term and a new product from Azure as well. We will be using AI Foundry today to build some of those initial agents as well. It might take few hours to deploy, but it gives you the real power. You can start connecting with your kind of a tools. You say, hey, I have a subscription from one place. Can I connect it? Yes, you can. I'm using this kind of a particular tool. I build my own small application. Can you connect with that? Yes, you can. Can I use my own database? Can I select my own specific things? You can. Finally, you say it's going to be a pro code. You can say a professional coder. You are an ML engineer, you are a developer. You can go any way whenever you want to. In that case, you might be seeing some frameworks like a land graph or a crew AI. These are some of the common themes, common frameworks that you started thinking it about. Multi-agent, they can connect with your graph workflow. Or if you're talking about different toolings, you can go for Anthropic MCP, that is a model contextual protocol, or an SDK. It's an open protocol, but you will be able to start connecting different kind of tools, different kind of softwares as a part of your multi-agent systems. Or you can go for a very custom build that will give you a full access, full authority, any model, any tooling, any kind of an application layer that you can do it. One interesting point is almost 80 percent of these enterprise agents are being currently built, or can it be easily built with no code. So when you say that, okay, I don't have a technical skills, I am not an engineer, there's no PhD required for this thing out. You can actually build your agent, and I'm going to show you how.