When should an AI agent decide on its own?

Autonomy is usually configured as a level — this agent is supervised, that one is not — and that framing does not survive contact with real work, because a single run contains decisions of wildly different consequence. The useful unit is the decision, and three properties determine whether it can be made without asking.

Reversibility is the first and the most important. A choice that can be undone in one command has a bounded cost even when it is wrong, and bounded cost is what makes delegation rational. This is why the practical route to more autonomy runs through version control, snapshots and soft deletion rather than through better judgement: making things reversible expands the set of decisions that can safely be delegated without changing anything about the agent.

Containment is the second. A decision whose effects stay inside the work — a structure, an ordering, an intermediate format, an approach — can be re-decided later at the cost of the work itself. A decision that leaves the work becomes a fact other things depend on: a record other systems read, a message someone received, a name that gets used elsewhere. The leaving is what makes it consequential, not the difficulty.

Information access is the third and the most frequently overlooked. An agent can only decide well from what it can see, and the decisions that go wrong are usually those requiring context that exists in somebody's head: which customer is sensitive, which record is authoritative, why the previous approach was abandoned. These look routine from the inside, which is exactly why they get made autonomously and exactly why they are wrong.

Applying the three gives a division most teams find surprisingly liberal. Nearly all execution decisions pass: how to structure the work, which order to do things in, what to call an intermediate file, how to phrase a draft, which approach to try first. Nearly all commitments fail: prices, dates, scope, anything sent, anything deleted, anything another party will rely on. The uncomfortable middle is small and it is where the design attention belongs.

One thing that should not appear in this calculation is how well the agent has performed recently. A run of good decisions is evidence about the decisions that came up, not about the ones that have not yet. Expanding autonomy is a reasonable thing to do as evidence accumulates, and the evidence that should move it is structural — a rollback now exists, a check now runs, the information is now written down — rather than a record of nothing having gone wrong.