What is an AI agent?

A chatbot answers. An agent does. That is the simplest way to tell them apart.

When you ask a chatbot for a flight, it writes a paragraph about flights. When you ask an agent, it opens a booking site, searches, compares prices, and comes back with a choice or a booking. The model underneath can be the same. What changes is the loop around it.

The loop

Every agent, from a coding assistant to a browser bot, runs some version of the same cycle:

1. Plan. The model reads the goal and decides on the next step.
2. Act. It calls a tool: a web search, a code runner, a calendar, a database, a browser.
3. Observe. The tool's result goes back into the model's context.
4. Repeat until the goal is met, a limit is hit, or it asks a person.

Researchers described this pattern in 2022 as "reasoning and acting": the model writes down what it thinks, takes an action, reads the result, and thinks again. Most products called agents today are built on that idea.

The parts

Where it breaks

• Small errors compound. If each step is right nine times out of ten, a ten-step task often fails somewhere. Long tasks are where agents struggle most.
• Tools return surprises. A page changes layout, a login appears, an API times out. The agent has to notice and recover.
• Instructions can hide in content. A web page or email can contain text written to steer the agent. This is called prompt injection, and it is the main security worry with agents that browse or read mail.
• Cost. Every loop is another model call. A task that takes fifty steps costs fifty times as much as one answer.

How to read an agent launch

Ask three questions:

1. What tools can it use? An agent is only as capable as the actions it is allowed to take.
2. When does it stop and ask? Good agents ask before spending money, sending messages or deleting things.
3. How long are its tasks? A demo of a five-step task says little about a fifty-step one.