Will it do something irreversible without asking me?

Two things stand between a run and an action you did not want, and it is worth knowing which is which.

The first is the agent's own judgement. Every agent step, in a learning run and in an autonomous workflow alike, is told to stop and ask before anything consequential: a message a person will read, money moving, data deleted, a form submitted, something posted in public. When it asks, the run suspends, the decision arrives in your inbox with a plain sentence describing exactly what it wants to do and why, and nothing proceeds until you answer. It is also told not to interrupt you for reading things or for work that is trivially undone.

The second is a gate you pinned. During the codification walk-through you mark the checkpoints where a person must always sign off, and those become a permanent part of the workflow. A gate fires on every run regardless of how confident anything is, and no amount of earned autonomy removes it. Pin them where the responsibility sits rather than where the machine looks unsure: the payment above a threshold, the letter to a regulator, the thing your name is on.

The difference is the honest part. The first is a model following good instructions, which works well and is not a guarantee. The second is enforced by the runtime, and it is. If a step must never go through without a human, pin it rather than trusting the agent to ask.

Either way the run waits rather than guessing, and it holds its place for as long as you need.