What Is an AI Agent? Start With What Came Before

An AI agent is an AI model that can take steps and use tools in a loop until a job is done. You give it a goal, it decides what to do first, does it, looks at the result, and picks the next step, stopping when the work is finished or when it needs you. Anthropic, which builds the Claude models, says in its engineering guide on building effective agents that agents "are typically just LLMs using tools based on environmental feedback in a loop." Tools, feedback, and a loop are the whole definition.

The word sounds like a leap, but you have probably climbed nearly all the way to it already. Search came first, then typing questions in plain words, then saved projects, then models working inside a folder on your computer. That climb has its own post, from search to projects to folders. The agent is the next step after the folder, the point where the model stops handing each move back to you.

What can an agent do that a chatbot can't?

Give a chatbot a stack of forty supplier invoices and it will tell you how to reconcile them. Give an agent the same stack in a folder, along with access to a spreadsheet, and it will open each invoice, pull the totals, enter them, notice that two invoices share a number, check whether one is a duplicate, and come back with the sheet and a short note about the duplicate.

Three things make that possible. The agent has tools, meaning specific abilities someone granted it, such as reading files, writing a spreadsheet, searching the web, or sending an email. It has a goal instead of a single question. And it runs in a loop, checking each result before choosing the next move. A chatbot does one step and hands the next one back to you, while an agent keeps the turn until the job is done.

What does the loop look like?

Slow the invoice job down and the loop has four beats that repeat.

1. Decide. The model reads the goal and what it has so far, and picks one next action, such as "open the next invoice."
2. Act. It calls a tool. The tool does the work outside the model, opening the file or writing the cell, and returns a result.
3. Check. The model reads what came back. Did the file open? Does the total match the line items? Is this invoice number already in the sheet?
4. Repeat or stop. If the goal isn't met, it goes back to deciding. If it is, or if the next action needs a person, it stops and reports.

The check is what separates an agent from a script. A script runs the same steps whether or not the last one worked. The agent reads each result, so when invoice 23 turns out to be a scanned photo instead of a PDF, it can try another way to read it or set it aside and say so. That flexibility is why agents can take on messier jobs than older automation could, and it's also why they need watching.

Where do tools come from?

An agent can only do what its tools allow. In a desktop app, the tools might be "read a file in this folder" and "write a file in this folder," and nothing else. To reach your calendar, inbox, or QuickBooks, the app needs a connector, a prebuilt link between the AI model and that other software. What MCP is, and how APIs and connectors fit explains that plumbing in plain terms.

The short version is that an agent can only touch what a connector opens, and only as far as you let it. A connector that reads your calendar gives the agent eyes. One that can also create appointments gives it hands. The order to grant those in is laid out in how much access to give an AI agent.

Where agents go wrong

The loop that lets an agent recover from a bad file also lets it repeat a bad idea quickly. If the goal is vague, the agent fills the gap with a guess and then builds forty steps on top of it. Ask one to "clean up the client folder" and it has to decide what "clean" means for you. A person would ask. An agent may ask, or it may start renaming.

Three habits keep that in check. Write the goal as one sentence with a finish line, such as "enter every invoice in this folder into the March sheet and list any duplicates." Keep the approval prompts on for anything that sends, deletes, or spends. And give the agent written instructions to work from, because the model only knows your business through what's on the page.

That last habit is the one I'd push hardest. My own instruction files started as a few lines each. They grew one correction at a time, every time I caught myself fixing the same mistake twice, and the corrections are now worth more to me than any single answer the model has written. When you correct an agent, ask it to write the lesson into the file so you never correct it again. Why files carry so much weight comes down to context, which the mechanisms behind the magic covers.

Is an agent the same as automation?

Not quite, and the difference decides whether you need one. Older automation follows a fixed path: When a form comes in, add a row. An agent chooses its path as it goes. Plenty of the jobs owners want off their plate are fixed-path jobs, and those run more predictably with no model in the middle. Which jobs are worth handing to software at all, agent or not, is the subject of AI automation for small business.

In the four levels of AI use, agents sit at Level 4, delegated, and the series' standing advice applies here as well. Use the lowest level that does the work.