When you see the word "autonomous" on a product page, what should you take it to mean? The answer is not clear, because the word is currently used for at least five very different things.
One company calls a plugin that suggests code autonomous. Another uses the same word for a system that opens and finishes work on its own overnight. Both are sold under one label, and the buyer only finds out what they bought after using it.
Why the car analogy falls short
The first fix that comes to mind is borrowing from the car industry. Driving automation has a six-level scale everyone knows, and "level 3" means something specific.
Carrying that logic over to agents is misleading. A car has one job, which is getting down the road without leaving it. An agent involves three separate questions, and a product can be far ahead on one of them and at the very start on another.
Three axes
The decision axis. Who picks the order of the steps? At one end the order is written entirely in code, at the other the model decides the next step at every step.
The authority axis. What can the system do? At one end it only reads and writes suggestions, at the other it updates records, sends messages, writes and ships code.
The trigger axis. Who starts the work? At one end there is a human at the start of every run, at the other a schedule or an outside event starts it.
Once you separate these three, the confusion in the market resolves. A plugin that suggests code sits high on the decision axis, low on authority and at the very bottom on trigger. A system that opens its own work overnight sits high on all three.
That is where the distance between two products sold with one word becomes visible.
Five practical levels
Combining the axes gives five rungs that most products land on.
Suggests. The model produces something and a human applies it. A wrong suggestion touches nothing, so the risk is low, and the gain stays limited because the load of the work is still on the human.
Asks at every step. The model acts but asks before each action. It works reasonably well while a single person is using it.
Finishes on its own within narrow authority. It completes the job inside a defined set of tools without asking, and the result comes to a human. Review happens after the run.
Carries its own trigger. No human starts the work. A schedule, an incoming event or a changed record starts the run.
Defines its own work. It is given a goal and decides which jobs to do. Very few products sold today are really here.
The hidden cost of the approval rung
Asking for approval at every step sounds like the safest option. In practice it produces a different result.
The first problem is that the gain melts away. A system that needs a person to wait at every step does not free up that person's time, and the whole point of automation disappears.
The second is approval fatigue. Someone who has seen twenty approval dialogs in a row clicks through the twenty-first without reading. At that point the approval step stops being a check and becomes a habit that thinks it is one.
We deliberately did not pick this rung ourselves. Instead of asking for approval mid-run we moved the safety earlier. The tools an agent can reach are narrowed up front, the spending ceiling is set before the run starts and every step taken can be read afterwards.
That choice has a price and there is no point hiding it. For work that cannot be undone, approval up front is still the right approach. Skipping it for a money transfer or an announcement going out to customers is not reasonable.
Raising the level is not always better
Autonomy is not a measure of success. It is a design choice, and the right level depends on how reversible the work is.
Where work can be undone, high autonomy pays off. A badly written draft gets deleted, a wrong analysis gets rerun.
Where work cannot be undone, low autonomy is the smarter choice. You cannot recall a message that has gone out.
So different jobs inside the same company should run at different levels. Picking one level and applying it everywhere leads either to needless waiting or to needless risk.
Four questions for the vendor
Does the model pick the order of the steps, or is the order written on your side? If the answer is "the model picks", you are high on the decision axis.
Which operations can the system carry out without approval, and how is that list narrowed?
Is a human the only thing that starts a run? If not, what else starts one and who can switch it off?
Once a run has finished, where do I see which steps were taken and what was spent?
The fourth question carries the most weight. Any system that removes approval up front has to put solid after-the-fact review in its place. A product that removes the approval and does not add the review is not autonomous. It is just running unsupervised.