What Are AI Agents?

Beyond answering: doing

A normal chatbot answers your question and stops. You ask, it replies, the loop ends. An AI agent is different: you give it a goal, and it figures out the steps to reach that goal — planning, taking actions, checking how they went, and adjusting. It keeps going until the job is done or it gets stuck.

The model inside an agent is often the same kind of language model behind a chatbot (see How Chatbots Work). The difference is the scaffolding around it that lets it act, not just talk.

A concrete example

Imagine you tell an agent: "Find three quiet cafés near the library that are open on Sunday and make me a short list."

A plain chatbot would guess from memory and might be out of date. An agent instead:

1. Plans — "I need to search for cafés, check their hours, filter by Sunday, then write a list."
2. Acts — it uses a search tool to look up cafés near the library.
3. Observes — it reads the results, noticing which list opening hours.
4. Acts again — it checks each café's Sunday hours, maybe using a maps tool.
5. Decides — it drops the ones that are closed or noisy.
6. Finishes — it writes you the short list.

Notice the rhythm: plan, act, observe, repeat. That cycle is the heart of every agent.

The agent loop

We can write the loop simply:

1. Look at the goal and what has happened so far.
2. Decide the next action.
3. Carry out that action using a tool.
4. Read the result.
5. Are we done? If not, go back to step 1.

This loop is what turns a one-shot text predictor into something that can tackle a multi-step task. Each trip around the loop, the agent learns a little more about the situation and gets closer to the goal — or realises it needs a different approach.

Tools: the agent's hands

On its own, a language model can only produce text. Tools give it real abilities. Common tools include:

• Web search — to find current information.
• A code runner — to do exact calculations or process data.
• File access — to read and write documents.
• Other apps' controls — to send a message, book something, or update a spreadsheet.

When an agent "uses a tool," it writes a structured request (like search: cafés near library), the system runs it, and the result is fed back into the loop. Tools are what let an agent escape the limits of the model's memory and act in the real, current world.

Assistants are agents you talk to

Many "AI assistants" are agents wrapped in a friendly chat. You make a request in plain language, and behind the scenes the assistant plans, calls tools, and reports back. The agent machinery is hidden; the helpfulness is what you see.

Why autonomy cuts both ways

Agents are exciting because they act — and risky for exactly the same reason. A chatbot that gives a wrong answer is annoying. An agent that takes a wrong action — deleting the wrong file, sending the wrong message, buying the wrong thing — causes real harm that can be hard to undo.

Good agent design takes this seriously:

• Limited tools. Only give the agent the powers the task needs.
• Human in the loop. Pause for approval before important or irreversible actions.
• Guardrails and logs. Set rules it cannot cross, and record what it did so people can review it.

The honest limits

An agent is only as reliable as the model inside it, plus its tools. So it inherits familiar flaws:

• It can hallucinate — invent a false fact — and then act on it.
• It can misread a tool result or get stuck repeating a useless step.
• It has no real understanding of consequences, so it will not "feel" that an action was a bad idea.
• Long chains of steps compound errors: a small slip early can derail the whole task.

This is why, today, the most trusted agents work with people rather than fully alone, especially for anything that matters. For the bigger picture of where this is heading, see The Future of AI and Humans.

Agents turn AI from a thing that talks into a thing that acts. That is genuinely powerful — and exactly why understanding the loop, the tools, and the limits matters before you hand one a real job.