Expert systems that operate in complex domains are continually confronted with the problem of deciding what to do next. Being able to reach a decision requires, in part, having the capacity to “reason” about a set of alternative actions. It has been argued that expert systems must reason from evidential information–i.e., uncertain, incomplete, and occasionally inaccurate information [Low82a]. As a consequence, a model for reasoning about control must be capable of performing several tasks: to combine the evidential information that is generically distinct and from disparate sources; to overcome minor inaccuracies in the evidential information that is needed to reach a decision; to reason about what additional evidential information is required; to explain the actions taken (based on such information) by the system. These are a few of the formidable control problems that remain largely unsolved [Bar82]. If expert systems are to improve their performance significantly, they must utilize increasingly sophisticated and general models for dealing with the evidential information required for reasoning about their behavior. To this end we present an alternative evidentially-based approach to reasoning about control that has several advantages over existing techniques. It enables us to reason from limited and imperfect information; to partition bodies of meta- and domain-knowledge into modular components; and to order potential actions flexibly by allowing any number of constraints (i.e., control strategies) to be imposed over a set of alternative actions. Furthermore, because it can be used for reasoning about the expenditure of additional resources to obtain the evidential information needed as a basis for choosing among alternatives, this approach can be employed recursively.