Incoming work units (e.g., requests) in a computing workload are analyzed and classified according to predicted execution. Preferred embodiments track which instrumented wait points are encountered by the executing work units, and this information is analyzed to dynamically and autonomically create one or more recognizers to programmatically recognize similar, subsequently-received work units. When a work unit is recognized, its execution behavior is then predicted. Execution resources are then allocated to the work units in view of these predictions. The recognizers may be autonomically evaluated or tuned, thereby adjusting to changing workload characteristics. The disclosed techniques may be used advantageously in application servers, message-processing software, and so forth.