Maybe the real limit on AI autonomy should be reversibility

I’ve been thinking about what actually makes me uncomfortable about letting an AI agent take actions.

It may not be the size of the action.

It may be whether we can undo it.

Consider two actions:

• An agent changes a $500 subscription plan, but the change can be reverted instantly.

• An agent sends the wrong email to 5,000 customers.

The second action may cost almost nothing directly.

But once the email is sent, there is no rollback button.

The same applies to things like:

• deleting data

• submitting a form externally

• sending a payment

• publishing content

• changing permissions

• contacting a customer

We often classify agent actions by risk, confidence, or monetary value.

But maybe we should also classify them by reversibility.

That could lead to a different way of designing agent workflows:

• Easily reversible → agent acts directly

• Reversible for a short period → agent acts, but execution is delayed

• Difficult to reverse → additional checks before committing

• Irreversible → human confirmation or a completely different workflow

And perhaps agents shouldn't just have permission to perform an action.

They should also know the rollback path before taking it.

If an action fails, what restores the previous state?

If the answer is "nothing," maybe that action deserves a very different level of autonomy.

For people building agents that take real-world actions:

Are you designing around reversibility today, or mostly around permissions and approvals?

And what actions in your systems turned out to be much harder to undo than you originally expected?