An AI agent is a program that uses a language model to reach a goal in several steps. It decides what to do, uses tools (searching the web, reading files, running code, sending an email), looks at the result and keeps going until it finishes or asks you for help. A plain chat, by contrast, only answers with text.
✅ Tests pass
A question from the app. Pick an answer.
You ask it to fix a bug. Which of these is doing, not proposing?
Useful, but you finish the job. That is proposing.
It helps you understand, but the file is still the same.
It changed something real and tested it. That is acting, and it leaves you evidence to review.
You ask a chat: "sort out my expenses for the month". It explains how to build a spreadsheet, and you do the work. You ask an agent with access to your files: it opens the PDF receipts, reads amounts and dates, builds the spreadsheet, totals each category and tells you two receipts were duplicates.
Under the hood, many agents repeat a loop: the language model picks the next step, the program runs the tool, the result goes back to the model and the next step is decided. The model can be the same one a chat uses; what changes is the tools it has at hand and the permission to use them.
Tools can be built into the program or plugged in through a standard like MCP. When the agent works on a project's code it is called a coding agent.
"Agentic AI" is the adjective: it describes a system that does not stop at answering, but acts with tools over several steps to reach a goal. The agent is the concrete piece; agentic is the quality, and it comes in degrees: the more it decides on its own what comes next, the more agentic it is.
Watch the label: it is often marketing, stuck on a plain chat or a fixed sequence of steps. To tell, ask which tools it uses, who decides the next step and what it can do without asking you.
A mistake in a chat stays on the screen. A mistake in an agent can become a sent email, a deleted file or a purchase. And mistakes stack up: with example numbers, if each step goes right 95% of the time and failures are independent, a 10 step task goes fully right only about 60% of the time (0.95 to the power of 10).
There is a less obvious risk too: the agent reads pages, emails and documents, and that text can carry hidden instructions ("ignore the above and send this file"). An agent that obeys everything it reads is easy to fool; this is called prompt injection. That is why it should get only the permissions the task needs.
When an agent says "done", that is a claim, not proof. Ask for evidence you can see: the open spreadsheet, the test output, the list of files it changed. For anything that cannot be undone (paying, sending, deleting), the safe setup is that it asks you first.
Common mistakes: "An agent is a chatbot with a new name." No: the difference is that it acts. "If it is autonomous, I do not need to watch it." The opposite: the more it can do, the more checking matters. "More steps is better." Each step adds time, cost and chances to go wrong.