What Is an AI Agent? A Simple Explanation

A word everyone uses #

“AI agent” is everywhere now. But many people use this word in different ways. Before I build my agents for CVE Friday Brief, I want to be clear about what an agent really is.

Here is my simple definition.

First, a plain LLM call #

An LLM (Large Language Model) is the engine behind tools like ChatGPT. On its own, it does one thing: you send text in, and it sends text back.

You ask a question. It gives an answer. Then it forgets everything. This is one call, one answer. Nothing more.

This is useful, but it is not an agent.

What turns an LLM into an agent #

An agent is an LLM that can decide its own steps to reach a goal. This is the key idea. With a plain LLM call, you control every step. With an agent, the model chooses what to do next, based on what it finds along the way.

To do this, an agent has three parts:

• A goal. Not one question, but a task to finish. For example: “Find this week’s most important CVEs.”
• Tools. The agent can do actions, not just talk. It can call an API, read a database, or run a search. (Tools are optional, but most useful agents have them.)
• Freedom to choose. The agent works step by step. It reasons, picks an action, looks at the result, and then decides the next step on its own. It repeats until the goal is done.

So the simple idea is:

An agent is an LLM that picks its own path to a goal, using tools when it needs them.

Agent or workflow? #

There is one more word you will hear a lot: workflow. It is easy to mix it up with an agent, so here is the simple difference.

• In a workflow, you decide the steps. The system always runs the same path. It is predictable.
• In an agent, the model decides the steps at runtime. The path can change every time. It is flexible, but less predictable.

Both are useful. In fact, most real systems mix the two: a fixed workflow as the “spine”, with agents handling the hard, flexible parts in the middle. I will come back to this choice in a later post.

A small example #

Imagine I ask a plain LLM: “Is Log4Shell (CVE-2021-44228) still a real threat?”

A plain LLM can only guess from old training data. It cannot check anything new.

An agent works in a different way:

1. It reads the goal: check the current status of this CVE.
2. It calls a tool to fetch live data from NIST NVD.
3. It calls another tool to check the CISA KEV list (known exploited vulnerabilities). Log4Shell is there - attackers still use it, years after it was found.
4. It looks at the results and writes a clear answer, based on real data.

Same question. Very different quality.

Why this matters for my project #

CVE Friday Brief is not one big question. It is a real task with many steps: collect data, choose what matters, add context, write a detection rule.

No single LLM call can do all of that well. But a small team of agents, each with its own goal and tools, can.

In the next post, I will show the full pipeline: how raw CVE data becomes a finished Friday brief.

Thanks for reading. The next post is coming soon - subscribe via RSS if you want to follow the build.