You have probably noticed that every AI product this year is suddenly an "agent". Agent for sales. Agent for support. Agentic browsers, agentic spreadsheets, agentic everything.
The word describes something specific. Here is what it means, in plain English.
A model on its own can only talk
Start with what ChatGPT or Claude is underneath: a model. It does exactly one thing. You give it text, and it gives you text back.
It cannot open your inbox, look at a spreadsheet, click a button, check today's date, or remember what you told it last Tuesday. A model can only work with what you carry to it, and you have to carry the answer back out yourself. The model is a very smart person in a locked room, passing notes under the door.
A model on its own
- Reads the text you paste in
- Writes a reply
- Forgets it by the next message
You are the arms and legs.
A model in an agent
- Searches, reads and queries for itself
- Takes a step, checks, decides the next one
- Works to a goal and knows when to stop
It fetches its own information.
An agent is a model that can act
An agent is what you get when you give that person in the locked room three things.
Tools. Ways to do things: search the web, read a file, query your calendar, draft an email, update a record in your CRM. Each tool is a specific, limited capability someone has wired up.
A loop. Instead of answering once and stopping, the model takes a step, looks at what happened, and decides what to do next. Run the search. Read the results. Notice they are stale. Search again with a better query.
A goal and a stopping point. Something to achieve, and a way to know it is done, including handing back to you when the next step is one a human should approve.
Agent = a model, in a loop, with tools, working towards a goal. Everything else is engineering around those four pieces.
Watch one work
Ask an agent: "Find the invoice from our printer supplier in June, check whether we paid it, and draft a chasing email if we didn't."
- 01Searches the inbox for invoices from the supplier in June. Finds one, reads the PDF.Read only
- 02Notes the amount and the invoice number.Read only
- 03Queries the accounting system for a payment matching either one.Read only
- 04Finds nothing, so checks the bank feed too, in case it was paid outside the system.Read only
- 05Still nothing. Drafts a chasing email and puts it in front of you.Needs you
Every step used a tool someone had connected. And the agent sent nothing: it stopped at the right moment and asked. A well-built agent knows which actions are cheap (reading, searching) and which ones need a person.
Why agents fail
When an agent can't do something, the reason is almost never that the model isn't clever enough. Some part of the machinery around it is missing.
| What you see | What is actually missing |
|---|---|
| It can't see your data | Nothing connected the data to it |
| It can't take an action | Nobody gave it that tool, or the tool doesn't allow it |
| It forgot your instructions | Nothing was set up to remember them between sessions |
| It did the task badly | It was never told how your company does that task |
Rephrasing the request does not create a capability that was never wired in. The fix is on the plumbing side: connect the data source, grant the tool, write the instructions down. That plumbing has names: connectors, skills, memory. Each one gets its own post.
What this means if you run a team
The gap between "we use ChatGPT a bit" and "AI does real work here" is not a smarter model. Everyone has access to the same models. The difference is whether anyone has connected the model to your actual work. That means your inbox, your documents, your systems, and your way of doing things.
That's a plumbing job, and it's very doable. It's also the work I do.