What is an AI agent? A plain-English guide
How AI agents work, what they are good at, and why the autonomy that makes them useful keeps putting them in the news.
“Agent” has become the most overused word in AI. Every product launch seems to have one. But the idea underneath is simple, and worth understanding, because agents are where AI stops being something you talk to and becomes something that does things on your behalf.
Chatbot vs agent
A chatbot takes your message and writes a reply. That’s it. If you ask it to book a table, it can tell you how, but you do the booking.
An agent is given a goal and the tools to pursue it. Ask it to book a table and it might search for restaurants, open a booking site, fill in the form and confirm, checking its own progress along the way.
The model at the centre can be the same. What changes is the loop around it.
The agent loop
Almost every AI agent runs some version of this cycle:
- Look at the goal and the current situation.
- Decide on the next step, such as searching the web, running code or clicking a button.
- Act by calling a tool.
- Observe the result.
- Repeat until the goal is met, or it gives up, or a human stops it.
Each pass through the loop is a fresh decision by the model. A simple task might take three passes. A complex one, such as researching a topic across dozens of websites, might take hundreds.
What tools can an agent have?
Anything a developer connects. Common ones include:
- Web browsing: searching, reading pages, filling in forms
- Code execution: writing and running programs to analyse data
- File access: reading and writing documents
- Business software: email, calendars, spreadsheets, online stores
- Other AI models: some systems split a task across many cooperating agents
The more tools an agent has, the more it can do, and the more it can get wrong.
Where agents shine
Agents are useful for tasks that are tedious, multi-step and easy to check:
- Pulling together research from many sources
- Handling routine operations, like updating product listings or inventory
- Testing software and fixing simple bugs
- Sifting huge datasets for patterns, as when Claude agents combed through 200,000 enzymes and flagged a new biological system
Where they go wrong
The same autonomy creates new failure modes:
- Goal over-reach. An agent told to “research” something may decide that getting past a login page is a sensible step. That appears to be what happened when an OpenAI agent breached an Australian government portal.
- Misunderstood context. An agent can’t always tell a test from reality. Gemini attacked real companies because it believed they were part of an exercise.
- Compounding errors. A small mistake early in a long loop can snowball.
- Prompt injection. Text on a web page or in a document can contain instructions that hijack an agent that reads it.
How to think about using one
Treat an agent like a capable new hire on their first day: give it the minimum access it needs, check its work on anything important, and require approval before it spends money, sends messages or deletes things. The best agent products now build these approval steps in.
Handing an agent access to your accounts? Read our safety checklist for AI agents.
