What is an AI agent?

A calculator runs one instruction and stops. A web server answers one request and stops. An agent is different. It works in a loop: it looks at what is happening, decides what to do, does it, and then looks again. It keeps going until the job is done.

An agent is a language model that runs in a loop and can take actions in the world.

The loop

Every agent, simple or complex, repeats the same four steps:

1. Perceive. Read what is happening. This could be a user message, an API response, or the result of the last action.
2. Reason. Decide what to do next. The model looks at what it sees and picks an action, or decides it is finished.
3. Act. Carry out the decision. Usually this means calling a tool.
4. Update. Save what happened so the next turn of the loop can use it.

The loop runs until the agent reaches its goal, hits a limit you set, or fails in a way it cannot recover from.

Why the loop matters

A single call to a language model returns text. That is useful, but the model cannot search the web, read a file, or run code. It can only produce words.

The loop changes that. On each turn the agent can call a tool and feed the result back in. Across several turns it can look something up, use what it found to do the next thing, and build toward an answer. Each step depends on the one before it.

What makes an agent good

Two things decide how capable an agent is:

• How well it reasons. Whether the model makes good choices at each step.
• How good its tools are. What the agent can actually do, and how reliably.

The rest of this course is about making both work in production: a clean loop, memory you can trust, failures you can recover from, and a clear view of what the agent is doing.

This course assumes you can already build a basic agent and want to make it production-ready. We use LangGraph throughout. The next lesson covers its two building blocks: tools and graphs.