The need for expressing and using metalevel knowledge is emerging in the design of several kinds of AI systems. The careful distinction between object-level and metalevel notions and the formalization of the latter has first been carried out by logicians for foundational reasons; subsequently, the distinction has been exploited in Artificial Intelligence and Computation Theory, revealing itself to be of great relevance to Automated Deduction and Problem Solving. This paper concentrates on the use of metaknowledge in building knowledge-based systems. In order to introduce the issue, some motivating examples are presented. We then review various paradigms for combining knowledge and metaknowledge, with the aim of abstracting general criteria that should underly the construction of viable AI systems, as far as metaknowledge is concerned. Furthermore, a general overview of the uses of metaknowledge in AI is provided and, among them, we concentrate on inference control, which can be conveniently exercised by formalizing control strategies at the metalevel and by letting the inference engine depend on metalevel descriptions. The technique is presented with the aid of some examples, chosen from practical AI applications, that are expressed in the formalism of Horn clause logic. The issue of self-descriptive systems is then addressed. A system that embodies and can use an adequate description of itself allows for self-evaluation (e.g., the estimate of the resources needed to perform a given task) and for self-modification (e.g., the automatic improvement of deduction performance by profiting from experience gained in previous deductions).