From Generative AI To Agentic AI
From the course: Agentic AI For Founders & Senior Operators
Generative AI produces words. Agentic AI produces words and then uses them. Part of that shift comes from agentic loops, where the AI carries context from earlier turns and checks its own work, and part from the agent harness, which lets ChatGPT and Claude browse the web and build decks instead of only chatting.
Transcript
First thing, what is the big shift in AI that has happened? A lot of us would have probably heard the term Generative AI and of late, probably say 2025, everybody's talking about Agentic AI and nobody's really had a framework for what shifted, why did we stop calling it generative, and why we stopped calling it agentic. At the absolute most fundamental, most basic level, the main thing that changed was that with Generative AI, we only used to produce words, right? And with Agentic AI, we now produce words and use them. There are a lot of things beneath that, that have driven this transition. So if you recall when ChatGPT first came out, it was single prompt and response. So it wouldn't even remember what you said in the previous prompt. And then this whole field of prompt engineering came up and people used to write these like perfectly crafted prompts. All of that is done. And why it's done is because of something called Agentic loops. So I'll show you what a loop is, but in a very, very quick sense, a loop is when an agent is able to, or an AI is able to use context from all the previous turns to perform the action it's trying to perform. And the second is that it checks its work as it does it, right? So it isn't a verbal vomit that comes out. It does something, pauses, checks, does something else, pauses, checks. So I'll show you a real-life example of how that's happening under the hood, which brings me to my second point that there was no self-editing with the Generative AI family of models. Now with Agentic AI, you would have heard a lot about these things called reasoning models and inference and such. That's just a fancy way of saying that it now spends more time thinking in the quote-unquote inference stage than it does in the training stage. In fact, if you look at it as a percentage of total compute, there has been a big transition from compute being used on training AI to AI using it for the purposes of inference, right? And inference is when it thinks, when it goes through the loop and so on. Third, ChatGPT, Claude only used to be able to chat. Today, probably all of us are using it to browse the web, create, you know, decks, et cetera, et cetera. I'll talk about how it's able to do that. It's because of this concept that was introduced called the Agent Harness. That's very important to understand because most of the changes that happen in AI today are actually happening at the harness level.