What is an AI agent? A plain-English guide

“AI agent” is one of the most overused phrases in tech right now. It’s used to describe everything from a fancy chatbot to a fully autonomous system that can ruin your weekend. So before deciding whether you need one, it’s worth being precise about what they actually are.

The shortest possible definition:

A chatbot answers. An agent does.

Let’s unpack that.

The chatbot vs the agent

A regular chatbot is a clever text machine. You ask, it answers. That’s the whole loop. Even when it sounds intelligent, it’s mostly producing words.

An agent does more. It can:

• Decide what step to take next.
• Use tools — call an API, run a search, query a database, send an email.
• Look at the result of that tool and decide whether to continue.
• Loop through several of these steps until the task is done (or it gives up).

An agent is built around an AI model, but the system around it is what makes it an agent. The model decides; the surrounding code actually does things.

The intern analogy

A useful mental picture: an agent is like an intern on day one.

You give the intern a goal — “book me a flight to Rome next week, under €300, departing in the morning”. The intern doesn’t already know the answer. They need to:

1. Open a flight-booking site.
2. Search for flights.
3. Filter by your constraints.
4. Compare results.
5. Maybe try a second site.
6. Come back to you with a recommendation, or do the booking.

That sequence of decisions, tools, and steps is exactly what an AI agent does — except the “tools” are software APIs and the “intern” is an AI model.

What agents are good at

Agents help most when the work has these properties:

• Multi-step. It takes more than one action to complete.
• Variable. Each instance of the task is a bit different — the agent has to adapt, not just follow a script.
• Bounded. There’s a clear definition of “done” or a clear way to stop.
• Recoverable. Mistakes are reversible, or there’s a human in the loop for the irreversible bits.

Good fits:

• Research and synthesis — gather information from multiple sources, summarise, cite.
• Triage and routing — read incoming messages, categorise them, decide who or what handles each.
• Data wrangling — pull information out of unstructured documents and put it into a structured form.
• Internal operations — checking statuses, generating reports, drafting communications that a human approves.

Where agents make things worse

Agents are not the right tool for:

• Simple, repetitive tasks. If the steps are always the same, you don’t need an agent — you need a normal automation. Cheaper, faster, more reliable.
• Tasks with irreversible side effects and no oversight. An agent shouldn’t autonomously transfer money, send to large customer lists, or make permanent changes to production systems without a human checkpoint.
• Real-time interactive UX. Agents are slow. Multi-step reasoning takes seconds to minutes. If your user is staring at a loading spinner, they’re not having a good time.

A common antipattern: a team replaces a perfectly working rule-based automation with an agent, because agents are exciting. The result is slower, more expensive, and less reliable.

Why agents fail

When agents go wrong, they tend to go wrong in characteristic ways:

• Loops. They take the same action over and over, especially when their tools return nothing useful.
• Hallucinated tool calls. They invent tools that don’t exist or call real tools with wrong arguments.
• Confident wrong answers. They produce a clean final answer that doesn’t actually reflect what they found in their tool calls.
• Wandering. They drift away from the original goal across many steps.

These aren’t reasons to avoid agents — they’re reasons to design them carefully, with step limits, schema-validated tools, self-checks, and evaluations that run before the agent is exposed to real users.

A pragmatic starting point

If you’re considering an agent, ask yourself three questions:

1. What’s the goal? Stated in one sentence, with a clear “done” condition.
2. What tools does it need? Listed explicitly — not “the internet.”
3. What can go wrong, and what stops it? Step budget, human checkpoints, what’s irreversible.

If you can’t answer those three questions cleanly, you don’t need an agent yet — you need to clarify the problem.

The honest summary

Agents are powerful when the work is genuinely multi-step and adaptive. They’re an expensive, unreliable replacement when it isn’t. The skill is knowing which is which — and building the system, not just the prompt, that makes the agent trustworthy.

Further reading:

• Building reliable AI agents — the failure modes and patterns that turn an unreliable agent into one you can deploy.
• Agentic RAG — what happens when agents and retrieval meet.