The Decoder/What is an AI agent, in plain terms?

What is an AI agent, in plain terms?

An AI agent is a language model that is allowed to take actions, not just write answers: it can open files, call tools, check its own results and keep going until a task is done or it gets stuck.

Mstimaj

By Fatima Jalloh • published 8 October 2026 • updated 30 September 2026 • free, no account

A chatbot answers one message and stops. An agent is given a goal, a set of tools and permission to use them, then works through the task in steps: look, act, check the result, decide the next move. It can search, read a file, run a command or send a request, and it loops until the job is finished. The model is the same kind of thing. What changes is what it is allowed to touch.

AI agent vs chatbot: what is the difference?

Take one task: book a call with a client next week. A chatbot can write the email that proposes times. You still open the calendar, find a gap, paste the times, send the email, and when the reply comes, add the meeting. An agent connected to your calendar and your mail does the loop itself. It reads the calendar, picks three open slots, drafts the email, sends it, waits for the reply, reads it, and creates the event. Six steps, each one a small decision that depends on the result of the one before.

That is the whole difference. A chatbot produces text. An agent produces text and then acts on it, with a tool, and looks at what came back.

AI agent vs automation: is it just a fancy Zapier?

Not quite, and the distinction matters for your money. An automation is a fixed recipe: when a form is submitted, add a row to the sheet, then send the email. Every run is identical, it costs pennies, and it never surprises you. An agent decides the steps as it goes. That flexibility is the point when the task is messy, such as sorting an inbox where every message is different. It is a liability when the task is regular, because the agent costs more per run, takes longer, and can wander.

A fair rule: if you can write the steps on an index card, build an automation. If the steps depend on what the last step found, that is a job for an agent.

What can an AI agent actually do today?

Whatever tools it has been given, and nothing else. The common ones are searching the web, reading and writing files on your computer, running commands in a terminal, calling a company’s API, and driving a browser. Coding agents such as Claude Code and Codex live in the terminal and treat your project folder as their workspace. Browser agents click through websites the way you would. Each tool is a door you opened. Close the door and the agent cannot go through it, which is also the safety model: an agent with no tools is just a chatbot again.

Do I need an AI agent for my small business?

Probably not yet for the routine work, and possibly yes for the one job that eats your afternoons. Routine and regular means an automation, which is cheaper and more predictable. Messy and judgement heavy means an agent, and it should start with one narrow job, on a read-only footing, with you checking its output for a couple of weeks before it is allowed to send or delete anything. The cost is not the subscription. It is the hour you spend deciding what it may touch.