What is an AI agent?
August 2026
Every AI product now says it comes with agents. Ask three vendors what the word means and you will get three different answers. That matters when you are the one paying, because the price of an agent and the price of a macro should not be the same number.
Here are the definitions we use in our own shop and with every owner we talk to. There are three layers, and the word "agent" only belongs to one of them.
Three layers, three words
- A tool does one thing when asked. Send this email. Look up this customer. Update these records. The AI can hold many tools, the way a tech carries many tools on the truck, but each one fires once, on request, and stops.
- A skill is a written procedure the AI follows. A recipe. "When we research a new lead, here are the nine steps, in this order, and here is what good output looks like." A skill does not act on its own. What it does is make every run of the procedure as good as your best person's run of it. Most of the value for a small business lives at this layer.
- An agent is given a goal, not a step. It plans the steps itself, uses tools and skills along the way, checks its own work, and comes back with a finished result for a person to review. The defining feature is that nobody sits between the steps.
The one-question test
Who takes the next turn?
Here is a pattern that shows up everywhere. Someone says they "hand work off to an agent." Watch what actually happens: a person feeds the AI a batch of data, reviews what comes back, feeds it the next batch, corrects it, and feeds it the next. The AI is doing real work, maybe through a live connection into the CRM. But a person is taking every next turn. That is a skilled operator driving a very good tool. Nothing wrong with it. It is often the right setup. It is also the reason the setup stops the moment that person goes to lunch.
If a person takes the next turn, you are looking at tools and skills. If the system takes its own next turn, and a person reviews the result at the end, you are looking at an agent.
Why the difference costs money
Two failure modes, and both are common.
The first: a vendor prices something as an agent when it is a macro with a chat window. You pay for autonomy you never receive, and someone on your staff quietly becomes the operator who takes every next turn.
The second is the reverse. An owner hears "agent" and pictures AI acting alone in their business, gets uneasy, and walks away from the whole category. But most of what that owner needs is tools and skills: faster lookups, procedures that run the same way every time, drafts that arrive ready for review. No autonomy required. Fear of the top layer keeps them from buying the two layers that were never scary.
Which one do you need?
- You need a tool when the job is a single action and a person should stay in the loop for each one.
- You need a skill when a procedure repeats and its quality depends on who runs it. Write it down once, and every run matches your best one.
- You need an agent when the whole loop can run without a person between the steps, there is a clear checkpoint where a person reviews the result, and someone owns keeping it running. That last part is real work. We wrote about it in The build is now the cheap part.
How we use all three ourselves
Our own shop runs on this stack. Tools connect our AI to email, the calendar, and our records. Skills are the written procedures we run every week, like the research checklist a new prospect goes through, versioned and improved like any other document. And we run true agents where the loop earns it. One example: before our own work ships, a reviewer agent attacks it. It works alone, from a fresh vantage point, and it keeps its own memory of the mistakes we tend to repeat, so it gets sharper with every review. Ask us about that one on a call. It is a good story, and before this post went live, it caught real problems in the post itself.
So when you hear "agent," ask the question: who takes the next turn? The answer tells you what is being sold, what it should cost, and what it will take to run.
Not sure which layer your problem needs?