Agentic AI. It’s the talk of the town and the phrase on everybody’s lips. Just uttering the spell in a boardroom has the ability to get everyone to open their checkbooks. In Europe alone, VCs have invested over $548 million in AI‑agent startups, and that’s in the first half of 2025 alone! This might have you asking: what is agentic AI and why is everyone so excited about it?
Large and in charge: LLMs
Our first step to understanding Agentic AI is to understand what made this all possible in the first place, that is the Large Language Model (or LLM for short). An LLM is a type of machine learning model that is able to understand the natural language we speak everyday.
For those who don’t want to be forced to do math by disembodied blog post, an LLM basically works by being trained on a large body of text (books, the internet, etc) and eventually learns to complete a sentence (given the start) by predicting the most likely words to follow.
To make LLMs more useful in real-world settings, researchers fine-tune them into chat or instruction models—versions that don’t just complete text, but follow commands like “summarize this” or “write an email.” This is what you would see in a simple chatbot or the ancient versions ChatGPT (3/3.5).
From Simple to Smart: Agents!
Traditionally, software engineers used LLMs as a small component in their programs. They would feed it text, get the LLM response and then the rest of the program would carry on and use that text in someway. This helped software engineers accomplish simple repetitive tasks with little effort, but that’s not what this article is about. What came next was the shake-up that had the entire industry chomping on the bit.
The real innovation came when researchers, programmers, and hobbiests began to realize these LLMs were suprisingly good at planning.This meant that an LLM can reasonably serve as the brain for an agent! Now here is what the all this groundwork has been leading upto: it’s time to explain, what is agent?
An agent is any actor(piece of software in our case) that can do the following things:
- An agent is able to observe the environment it exists in
- An agent is able to make a plan to achieve its goal
- An agent is able to interact with its environment until its goal is met
Any agent that uses an LLM as its brain to think and plan can be called Agentic AI. Agents are given tools to read and interact with their environments to see if their plan is working. Modern day ChatGPT is an example of agentic AI. If you ask it “what is the weather in Boston today?” The LLM will think that it needs to lookup the weather to achieve its goal of answering your question, and chatGPT will use a tool to lookup the realtime weather in boston. This info is given back to the LLM and the LLM will realize it’s goal is accomplished and it will generate your answer.
Pros and Cons
Now that we understand Agentic AI, lets talk about why everyone wants one! Agents are able to accomplish many high-level tasks without the need for manually programming each and every task. An agentic chatbot can tell you the weather, tell you a joke, debug your code, all without the need for explicitly programming these things. Infact, given the right tools and a smart enough brain, agents can accomplish anything. What’s even better is that these goals can be set as natural language meaning anyone can work with them, not just the tech community
So does that mean we should make everything agentic? ofcourse not! There are many reasons why you would want to avoid using Agentic AI in your systems. Firstly, its expensive, both to you and the environment. For reasons outside the scope of this blog post, using an LLM consumes a lot of power. This intensive power consumption drives up prices and increases emmisions. Moreover, agents have to understand a goal, come up with a plan, use tools and see if they worked, then return an answer. If you compare that to programs that accomplish the same tasks using defined algorithms, you’d notice the slowdown immediately. Finally one must consider that giving an AI the ability to make decisions might not be suitable for all cases (for ethical and reproducibility reasons).
Given the right set of tasks and enough time, agentic AI can produce substantial benefits and revolutinize the way we code and interact with our software. Its spurred seismic shifts in the tech industry and for good reason, but before you rush to throw agentic ai into your code base be sure to ask yourself if it’s truly the right tool in your kit for this situation. Happy coding!