Understand

Autonomous AI agent

An autonomous AI agent is a program that is given an objective, works out the steps to reach it on its own, uses real tools, observes the outcome of its actions and starts over until it is done. What separates it from a conversational assistant is neither the quality of its language nor the size of the model behind it: it is the right to act without being told every move, and the duty to check what it has produced.

1. Definition

An autonomous AI agent is software built around a language model, entrusted with a goal rather than a procedure. You do not describe the sequence of operations to perform: you say what you expect, you give it access to the tools it needs, and you let it work out how to get there.

The word “autonomous” invites a misunderstanding best cleared up straight away. It does not mean “beyond all control”, nor “deciding its own objectives”. An autonomous agent stays bounded by the mission it was given, by the tools it was opened up to and by the limits it was set. Autonomy applies to the route, never to the destination.

The practical distinction is simple: ask yourself what happens when the situation is not the one that was expected. A conventional program stops or produces a wrong result. An agent notices the gap, looks for another way, and reports if it finds none. That capacity to recover — and it alone — is what justifies the term.

2. The four parts of an agent

An agent is not a single part but an assembly. Remove just one and it goes back to being an assistant that answers.

An objective. An instruction that describes a state to reach, not a list of orders. “Make sure no invoice goes more than thirty days overdue without the customer being chased” is an objective. “Send an email to customer X” is a command.
Tools. What lets it act on the world rather than talk about it: reading a mailbox, writing to a database, calling a service, producing a file, browsing the web. A model without tools can only produce text, and therefore nothing that changes the company.
A memory. What it knows about its job, about the company and about what it has already done. Without it, it starts from scratch every morning, chases the same customer twice and forgets the rule it was reminded of yesterday.
A loop. The mechanism that makes it observe the outcome of its action, compare it with its objective, and decide whether to continue, correct or stop. It is the most inconspicuous piece and the most decisive one: without a loop there is no agent, only a slightly elaborate model call.

The loop deserves a further word, because it is what most demonstrations skip over. An agent that announces a plan and then stops has not looped. An agent that carries out its plan without ever looking at what it produced has not looped either — it has simply gone fast, and will be wrong fast. The full cycle — perceive, decide, act, learn — is what turns a model into a worker.

3. The five degrees of autonomy

“Autonomous” is not a switch but a scale, and confusion between its rungs explains most disappointments. A company buys degree 4 while picturing degree 2, or the other way round.

Degree 1 — suggest. The agent proposes, the human does. A draft written, a reply suggested. No risk, and no real time saved either: the work is still there in full.
Degree 2 — prepare. The agent produces the finished deliverable, and the human only has to approve and send. This is the most profitable degree to start with, because mistakes are visible before they leave the company.
Degree 3 — act below a threshold. The agent acts alone below a declared limit and asks above it. Chasing a €200 unpaid invoice on its own, seeking approval beyond €5,000. This is the cruising regime of a well-tuned system.
Degree 4 — act and account for it. The agent carries the task through from end to end and leaves a readable trail. Suited to reversible, repetitive work: filing, sorting, keeping things up to date, publishing internal monitoring.
Degree 5 — direct other agents. The agent breaks an objective down, distributes it, tracks progress and arbitrates. This degree only makes sense once the previous ones have been proven — handing it to a system that has never been measured amounts to multiplying an unknown error.

The degree is not chosen once and for all: it is chosen per task. The same agent can act alone when chasing a quote and ask a human before granting a discount. A system that can only set autonomy globally forces a choice between the useless and the reckless.

4. What an autonomous agent is not

It is not a chatbot. A chatbot waits for a question and returns an answer; the work starts after it, and a human does it. An agent is judged on a result in the real world: the invoice has gone out, the file is filed, the meeting is booked.

It is not an automation script or RPA. An automation script follows a sequence of steps written in advance and breaks at the first deviation — a column heading that changes, a screen that is redesigned. An agent pursues an objective and absorbs the deviation. The difference shows in the maintenance cost, not on the day it goes live.

It is not a copilot. A copilot speeds up the person at the keyboard; take the person away and nothing happens. An agent works when nobody is watching — which is precisely its value, and precisely why its work must be reviewable.

It is not general intelligence. An agent has neither intentions of its own nor any wish to do anything other than its mission. Fears of a system that would “take control” miss the point: the real risk, mundane and frequent, is an agent correctly carrying out a badly worded instruction, at scale, without anyone noticing.

5. What makes an agent reliable

An agent's reliability owes very little to the model driving it and a great deal to four safeguards. Companies that give up after three months almost never had a model problem.

A declared boundary. The list of actions that require human approval before execution — committing money, signing, replying to a sensitive customer, deleting. Declared, not left to the judgement of the moment.
An enforceable budget. The spend is reserved before the model is called, not recorded afterwards. A ceiling checked after the fact is not a ceiling, it is a report.
A readable trail. The objective pursued, the information consulted, the action taken, its outcome, its cost. Delegated work that cannot be reread becomes work that is no longer trusted — and the company quietly goes back to doing it by hand.
Least privilege. Each agent receives only the tools its mission requires. A monitoring agent has no business with write access to billing.

These four points are not comfort options: they decide whether you will be able to delegate more next year. That is the subject of the page on governance and budget control.

6. Six questions that expose a fake agent

Many tools present themselves as agents when they are really chaining model calls along a fixed rail. Six questions are enough to settle it, and they can be asked during a demo.

What does it do when a tool fails? If it stops dead, it is a chain. If it tries another way and then reports, it is an agent.
Can it repeat the same step twice? A real agent goes back and starts over. A rail only ever moves forward.
Does it know what it did yesterday? Without memory, it will chase the same customer twice and forget your correction.
Where does it stop on its own? If there is no action it refuses to carry out without approval, the boundary is not declared — it is absent.
What did this task cost? If the figure is not available per task, no judgement is possible about what is worth delegating.
Can I reread its work without asking someone? If the answer is no, control rests on trust, and trust does not scale.

The sixth is the one that settles matters fastest in practice. Being able to see what was done, by whom and at what cost is not a decorative dashboard: it is what tells you whether the system is working or merely busy.

7. Frequently asked questions

What is the difference between an AI agent and an AI assistant?

The assistant answers a request and hands back control; you are the one who acts. The agent pursues an objective, uses tools to act, checks its result and starts over if need be. The shortest test: if the tool produces text that someone then has to copy elsewhere by hand, it is an assistant.

Can an autonomous AI agent act without supervision?

On reversible, repetitive work, yes — that is the whole point. On anything that commits the company towards a third party, no, and not for a technical reason: because mistakes there are costly and sometimes irreversible. A serious system makes that boundary explicit instead of leaving it to everyone's caution, and asks for approval before acting rather than reporting it afterwards.

Which model do you need to run an agent?

The choice of model matters less than people think once a reasonable quality threshold is passed. What decides the outcome, in order: the clarity of the instruction, the quality of the tools provided, the memory of the job, and only then the model. A vague instruction cannot be rescued by any model, and a more expensive model will make the same mistake faster.

How many agents should you start with?

One, on a task that comes up several times a week, whose rule fits in a single sentence and whose result can be checked in seconds. Teams of agents make sense afterwards, once the first one has shown what works. Starting too big is the most frequent and most expensive mistake.

Can an autonomous AI agent get the objective wrong?

It does not set itself a new one, but it can read yours differently from how you meant it — especially if the instruction carries an unspoken assumption that everyone in the company shares without ever having written it down. That is why the first indicator to track is human rework time: if it does not fall, it is not the model that needs changing, it is the instruction that needs rewording.