What is an AI agent?

An AI agent is a language model that can take actions, observe what happened, and decide what to do next — repeatedly, without a human approving each step.

The word has been stretched to cover almost every AI product on the market. It's worth being precise, because the difference between the rungs below is the difference between a demo and something you'd trust with your calendar.

The ladder

Rung	What it does	Is it an agent?
Chatbot	Text in, text out	No
Retrieval chatbot	Looks things up before answering (RAG)	No — retrieval isn't action
Tool use	Calls a function you defined, once, returns the result	Barely — no loop
Agent	Chooses tools, acts, reads the outcome, decides again, loops until done	Yes
Multi-agent	Several agents delegating to each other	Yes, and harder to debug

The dividing line is the loop. A model that calls a weather API once is using a tool. A model that checks the weather, notices the flight is cancelled, searches for alternatives, and rebooks — deciding each step from what it just learned — is an agent.

Where the idea came from

The pattern was formalised in 2022 as interleaving reasoning with acting: let the model think about what to do, do it, observe the result, and think again. Reasoning alone produces plans disconnected from reality. Acting alone produces flailing. Alternating them produces something that can recover from surprises — which is what makes it useful, and what makes it hard.

What actually makes them hard

Not the loop. The loop is easy. The problems are:

Errors compound. A 95%-reliable step is about 36% reliable after twenty of them. Long chains fail as a matter of arithmetic.
Nothing knows when to stop. Agents get stuck retrying, or declare victory early.
Actions are real. A wrong sentence is embarrassing. A wrong DELETE is an incident. Reversibility matters more than accuracy.
Cost is unpredictable. Every loop iteration is a full model call, re-reading the whole context. Runaway loops burn real money.
Debugging is archaeology. When a twelve-step run goes wrong at step four, you're reading transcripts.

What this means for you

Ignore the word on the box and ask two questions: what can it actually do without asking me, and what happens when it's wrong?

The useful agent deployments today are narrow, reversible and observable — a coding agent whose work you review as a diff, a research agent that returns links you can check. The pitch that it will independently run a workflow end-to-end while you sleep is, for now, mostly a pitch.

What we addedA capability ladder separating what genuinely counts as an agent from what is a chatbot with a marketing budget.

This article was researched and drafted with AI assistance from the sources listed below, then checked and edited by Fiqhro Dedhen before publication. How we work.