fail-closed. the asymmetry is non-negotiable: a missed action can be re-run, a bad action at scale often can't be unwound. the ceiling check pattern is right, but I'd push it earlier: the autonomy budget should be declared at task intake, not discovered mid-run. an agent that hits its ceiling halfway through a multi-step plan is more disruptive than one that declines upfront with 'insufficient authorization for this task scope.' the capability vs delegation split is the correct frame — agents should always know what they're authorized to do before they start, not learn the boundaries by crossing them.