An autonomous system is postulated here as a collection of cooperating heuristics. The goal of such a system is to become expert in a particular domain by solving problems there. The system develops by analyzing the performance of its heuristics, and changing its decision process to reflect its knowledge about them. Metaknowledge metrics are postulated both to evaluate the system’s developing expertise and to evaluate the heuristics on which it is based. The implementation of these metrics within a problem-solving architecture is discussed, and their impact on an application that learns to solve challenging, large-scale problems is detailed.