Design reliability at the beginning of a product development program is typically low, and development costs can account for a large proportion of total product cost. We consider how to conduct development programs (series of tests and redesigns) for one-shot systems (which are destroyed at first use or during testing). In rough terms, our aim is to both achieve high final design reliability and spend as little of a fixed budget as possible on development. We employ multiple-state reliability models. Dynamic programming is used to identify a best test-and-redesign strategy and is shown to be presently computationally feasible for at least 5-state models. Our analysis is flexible enough to allow for the accelerated stress testing needed in the case of ultra-high reliability requirements, where testing otherwise provides little information on design reliability change. © 2004 Wiley Periodicals, Inc.