Improving AI safety by restricting what actions an agent can take, rather than trying to detect bad behavior after it happens.