An AI agent acting for you may be told to check with you before doing anything irreversible. That only works if it knows what irreversible means to you. Aarik Gulaya, who built Base Layer, ties the definition to tolerance:
For a definition of irreversibility, that is directly related to an individual, an org's or provider's or government's tolerance.
It depends on the person
The same act can be minor for one person and serious for another:
One person could say an irreversible act is scheduling a call at an inconvenient time, whereas another person may say sending payment to an unauthorized payee
So a fixed list of irreversible acts will not fit everyone. What counts depends on whose tolerance the agent is working within: an individual's, an organization's, a provider's or a government's.
What is still open
Knowing whose tolerance applies does not settle how an agent learns it. He asks whether intent can be inferred or has to be given, and expects the answer to differ from person to person:
... If there's something where there's potential irreversibility or the evolution of something after a particular date, can the intent be inferred, or has it been given in order to make a reasonable decision? But ultimately, this would vary from individual to individual.