What Is an AI Agent and Why Should You Care?

The name's Intelligence. Artificial Intelligence.

If I'm a dad, I can get away with bad jokes like that, right?

But really, what do agents and AI have to do with each other? Don't tell me we're going all James Bond on this.

While I can only imagine what MI6 is doing with AI, when we talk about agents in this space, we're most assuredly not talking about our favorite British secret agent.

So, what are we talking about?

To answer that question, we first need to understand a bit about Large Language Models as they currently stand.

In my last post, I talked about LLM's and introduced some simple use cases for them. So, what is an LLM? An LLM is a natural language processing powerhouse that's able to consider amazing amounts of context and data. You can talk to it like another human and it will respond in very human-like ways. It can provide detailed answers about many different topics and is a great tool for you if you need to generate content or brainstorm about an idea.

But LLMs are limited in what they can. At the end of the day, they are only a language model meaning they are very good at questions and answers but are not so good at querying a database or booking a flight or telling you the weather.

In fact, they're not even good at counting. If you ask an LLM "How many times does the letter a appear in this sentence?" you'll get some surprising answers.

But wait, LLMs are amazing, right? Why do they get basic things wrong? Isn't there something we can do to make LLMs better at these things?

Glad you asked.

That's where agents come in.

Agents are LLMs that use various models or tool to reason about and accomplish the task the user is requesting. This means that you could have one initial LLM that processes requests, determines the best LLM or tool to use to accomplish the task, then hands off the request to those systems. Or it could mean you equip one LLM with a number of tools to accomplish various tasks. Really the configuration of this could take ifinite different shapes, but the core of agentic LLMs is that you have multiple LLMs and/or tools working together to accomplish a task.

So... that's cool... but why should you care?

Agents in AI make LLMs much more powerful. Not only can they correctly count the number of a's that appear in this sentence, they can tell you the current weather, provide you summaries of the latest news, book a flight, get an Uber, order dinner, or potentially act on your behalf in your favorite apps.

How's this possible?

This is all very new, so it's mostly developers that are setting these things up right now and figuring out the possibilities, but here's a high level of how this would work.

Let's say you have an agent that you like. Perhaps it's Claude or Llama or another. Let's say you have it running on your computer. Now let's say you often travel for work. You don't love booking flights but it's a necessary part of your job. One day you learn that your favorite airline has released a tool for you to connect your LLM to (we call these MCPs because the implement the Model Context Protocol, but, more broadly they're called tools or agents). You connect your LLM to their MCP, authenticate, and decide to try it out.

"Book me a flight to San Fransisco leaving on May 15th and returning on May 18th," you ask your LLM.

You see that the LLM is processing your request, identifies that it has a tool to book flights, and commences to book the flight.

You'll likely need to provide a few more details and answer a few questions so it can complete the order, but when you're done with that, you get an email confirming the flight was booked.

And you never went to the website once.

All you did was ask your LLM to do this for you and, through the tool you connected to it, it did it for you.

Booking flights is only one example. What if you could shop from your LLM? What if you could book appointments, pay bills, consolidate accounts, trade stock, or perform some vital job functions? Agents make this possible.

Another use case of agents is having multiple LLMs. If you have a few, specialized use cases, you may want to develop a few, specialized models to handle those cases. You may have one general model for triaging user requests, then you may have a customer support model to handle customer issues, a sales model to intake prospective customers, or an education model to handle training needs. The triage model would intake the request then determine which model will best service the requests. Once it does that, it can either act as mediator between the user and the other LLM or it can dispatch the request to the other LLM and let it take over.

If you use the latest models, you'll notice they are starting to behave in this way. After you enter a prompt, you'll see that it analyzes your prompt then determines the best way to handle it. Some of these LLMs have tools built in to do various functions like counting. You'll see it uses those tools to handle your request and give ou the best answer it can.

So, when do you need agents?

Typically agents are used when the steps to accomplish a task are not clear. The LLM will need to reason about how to accomplish the task. Agents are also used to give the LLM hands enabling it to act on your behalf and do things LLMs typically cannot do such as book a flight.

So, why should you care? While LLMs revolutionized AI and brought its abilities to the people, Agentic AI is bringing the tools to the LLMs enabling them to do things they previously were unable to do. This has the potential to completely change the way you interact with apps and websites. Instead of going to the website, you'll talk with your AI and ask it to do things for you.

Basically, you're making your own personal assistant. That's right. You could have your own version of a Jarvis like Iron Man. Just likely not with all the holographic displays.

Agentic AI is causing huge waves in the AI space right now. It'll be interesting to see where it goes, but this much is clear: it's here to stay.