What Is an AI Agent? Loops, Tools and Examples

Agents are not chat windows

A chatbot produces one reply and stops. An agent runs a loop: it observes, decides what to do, calls a tool, reads the result, and repeats until the task is done. That loop — not the interface — is what makes an agent.

If you can name the loop, you can recognise an agent. If you cannot, it is probably a chatbot with a long prompt.

The decide-act loop

• Receive a goal (the user's request).
• Decide what to do next (reason about the goal and current context).
• Act: call a tool (search, code interpreter, database, API) or produce output.
• Observe the tool's result and feed it back into the loop.
• Repeat until the goal is met or a stopping condition triggers.

This is the same pattern the Agentic AI Engineer program builds on, module by module.

The three components

Every real agent has these three parts, whether or not the marketing calls them out:

• A model that makes the decisions.
• Tools it can call to affect the world or gather information.
• A loop that connects them and decides when to stop.

Memory is a fourth component that agents often need, because a task that takes ten steps requires remembering the first three. But it is an addition, not a requirement — a single-step agent is still an agent.

What makes it "intelligent"

• The agent can choose between several tools.
• It can recover from a failed tool call.
• It keeps enough memory of earlier steps to make consistent decisions.

The middle one is the real test. A system that runs a fixed sequence regardless of results is an automation script with an LLM stapled to the front. A system that notices a call failed and changes strategy is an agent.

Simple example

A support agent reads an email, searches your product docs, drafts a fix, and — only if the change is reversible — applies it. Each of those is one decision in the loop.

Notice the last step: it checks whether the change is reversible before acting. That is not caution from the developer, it is a design requirement for anything that touches the outside world.

Distinguishing agents from automation

Automation runs a fixed script. An agent adapts its next step based on what just happened. The difference is feedback: automation has none, agents react to results.

A practical test: if you can describe exactly what the system will do, in order, for any input, it is automation. If you cannot, and the path depends on what comes back, it is an agent.

Where the term gets overused

"Agentic" has become a marketing word attached to products that are really just a chatbot plus one function call. When someone claims a system is agentic, ask three questions: does it call more than one tool, does it decide between them, and can you see what it did? If the answer to any is no, it is a chatbot with extra steps.

Go deeper

• Contrast agents with plain chatbots in agentic AI vs chatbots.
• Learn the fundamentals of LLMs first with how LLMs really work.
• Build your first loop by hand using the first agent tutorial.