With the one I’ve made, it’s best effort. We record whether or not tool calls have succeeded and if the agent dies without knowing whether or not the call succeeded, then when we resurrect it we tell it the state is unknown and it can either inspect the resource or try again depending on the side effects.