What Is An AI Agent?

AI Agents, Agentic AI…maybe like me you have also been hearing a lot about this recently.

Sometimes it's in scary headlines like "AI Agents coming for your job!" or "I swapped my entire workforce for AI agents and moved our headquarters to the Death Star".

But what exactly are AI Agents and are they really coming for your job?

Let's dive in:

An AI agent is a software system that uses AI to make decisions and take autonomous actions in pursuit of a defined goal — without requiring step-by-step human instruction for each action it performs.

Or in other words… an AI Agent is AI which acts, not just AI which answers your questions.

Actually that is what the "Agent" part of the expression is talking about. Not an "Agent" like James Bond, but rather it describes the agency that the AI has.

You give it permission to do stuff without asking you (aka: agency).

AI Agents are considered to be pretty cool because they have changed the way we interact with AI to get stuff done.

Previously, we would interact directly with the AI model, maybe giving it a set of scenarios and asking it to do a task for us based on this.

This would be a "passive" AI model - it waits to be told what to do.

AI Agents are a bit more confident. They do stuff without being told to do it.

Kind of like having a really good personal assistant who quietly sorts out your calendar and diary, without you having to actually tell them to do it.

This is "reactive". Responding to a situation without being told exactly what to do.

AI Agents are also "always listening". They work in continuous loops, listening for changes in their environment and deciding when to react.

This means that they can do things again and again in a loop, unlike a single prompt/response solution, which is only executed once.

The upside to this is that we can get things done with limited or no human interaction, meaning that that lovely human has time to maybe play with their kids, watch a sunrise or do any other stuff that is infinitely better than work.

A summary of key AI Agent characteristics:

OK sounds pretty good, who doesn't love to watch a sunrise every now and then.

But how do AI Agents know what to do?

First of all, the AI Agent needs a goal to work towards.

(don't we all?)

Before the Agent can be set to work, we need to clearly define what we want it to do. For example:

Once we have defined this, the AI Agent can then work to complete this goal by doing the following steps (we will look at each of these in turn):

It repeats this loop until the goal is achieved or it can't progress any further (an error perhaps).

Let's look at each of these steps in turn:

Think of an AI Agent like one of those submarine radars in spy films.

They are constantly scanning their environment, taking in every detail, watching for any blip on the radar, any change in the current state.

This "radar" might look different depending on what you need your AI Agent to do.

For example, it might mean constantly scanning your inbox, looking for spam mail to sort. Or maybe constantly scanning your sales database, waiting for a refund to pop up that needs to be processed.

Of course, in order to observe its environment, we need to provide the AI Agent with access to the tools and services that it might need in order to do this.

This might mean connecting the AI Agent with external tools like Outlook/Gmail or your database. Or maybe giving the Agent access to the internet so that it can search and scan for information in real time.

The important part here is:

Connect the relevant tools and information so that the AI Agent has constant access to its relevant environment.

Quick side note:

How does the Agent know what tools are available? — give it up for MCP!

Remember MCP? This is exactly where it fits in. The agent asks the MCP server "what can you do?" and the server responds with a list of available tools and what each one does. The agent then knows what actions are possible and can choose between them intelligently.

(Check out my post on MCP if this is absolute gibberish to you)

Ah yes. Thinking. Usually a good idea for most situations.

Now that the submarine radar has detected a little "blip" on its radar, it needs to do something!

Maybe make a loud noise? Call the president?

The second step for any AI Agent is to decide what it should do based on a perceived change in its environment.

This is where the "AI" part of "AI Agent" comes in.

We use an AI model for the thinking part of the process. The model uses the relevant information from Step 1 to decide what action it is going to take.

For example:

The "think" part is what makes AI Agents different to standard automation workflows.

Automation workflows work by just doing things one after another, always the same, never adapting or changing.

AI Agents get to decide for themselves what to do.

Pretty powerful stuff!

Once the decision has been made, the AI Agent needs to act on it.

This is where those connections to tools come in really handy.

Thanks to the connection with your email account, the AI Agent has permission to send an email straight to the client.

Because we connected our database to the Agent, it can directly mark the refund as "processed" once it has been issued.

And the cycle continues!

A property of AI Agents is that they work in a continuous loop. Just like how our submarine radar doesn't just switch itself off after it spots the first nuclear bomb heading towards it.

That would be a pretty crappy radar.

The AI Agent, having performed its task, quietly waits in the background, listening for further changes in the environment and for the cycle to start again.

AI Agents sound pretty useful! So how can I use one in my day-to-day life?

There are fortunately lots of different ways that anyone can start using AI Agents in their real life, depending on how complex your "goal" is:

Great for simple use-cases or relatively low complexity goals.

Tools like Claude and ChatGPT already have agents built in to them. You can just describe a multi-step task and they'll figure out the steps themselves.

This is great for trying out simple agents and getting your hands dirty, without writing any code yourself.

If your goal is a little more complicated or you just want to keep costs down by using an open source tool, then n8n is a good way to go.

(Check out my post on n8n if you are not sure what this is)

With n8n, you can build more complicated AI Agents without the need to code, and provide your Agent with access to a larger range of tools. n8n then deploys your Agent so that it can run in the background all the time.

If you are handy with code, you can of course build and deploy your own AI Agent by writing the code yourself.

Slightly more complicated but great for specific use cases or when you have multi-agent systems (more than one AI Agent) working together.

My last note on AI Agent would be, don’t let people confuse you into thinking it’s all that deep.

AI Agents are just bit of code, connected to AI models and tools, deployed online which are constantly running and doing useful things for you.

The most important bits from this post: