Auditing what an agent did is only half the job. Companies also have to control what agents may do, and make those limits hold when systems fail.