What is an AI agent?

In plain English

A chatbot responds. An agent acts. Given a goal, an AI agent breaks it into steps, decides which tools to use — a web search, a calendar, a spreadsheet, a company system — carries out the steps, checks the results and adjusts. The language model provides the reasoning; the surrounding software gives it hands, in the form of permissions to click, search, write files or send messages.

The practical difference is the loop. Instead of one question and one answer, an agent keeps going through several rounds until it decides the job is done or it gets stuck.

Why it matters

Agents shift AI from drafting text to completing tasks, which is where most of the claimed business value lies — and also most of the risk. They also change the question people need to ask about any AI product: not just “how good is the answer?” but “what is it allowed to do, and who is accountable when it does the wrong thing?”

An example

You ask an assistant to arrange a team lunch. It checks four colleagues’ calendars for a free slot, searches for nearby restaurants that cater for a nut allergy, books a table, and emails everyone the details — several actions, no further instruction from you.

What to watch out for

Errors compound. A model that is right 95% of the time on a single step can fail badly over a twenty-step chain, and mistakes made with real permissions — a booking, an email, a deleted file — are harder to undo than a bad paragraph. Agents can also be misled by instructions hidden in the web pages and documents they read. Start with narrow permissions and a human approving anything irreversible.