Participants switched frequently between high/low and odd/even classification of a digit. The interval between a task cue and the next digit varied between blocks. In Experiment 1, the task switched predictably every two, four, or eight trials. In Experiment 2, switching predictably every four trials was compared with random switching. With predictable switching, the cost was limited to the first trial of a run. Random switching produced a more gradual approach to asymptotic performance. After one performance, control processes attenuate the resulting change in task-set bias if a further switch is likely, but this strategic modulation is soon overwhelmed by task-set priming through further performances. Preparation reduced switch costs but not interference from the irrelevant attribute: Control of interference appears to be reactive, not proactive. Switch costs did not increase with run length, suggesting that retrieval of the task set last associated with the stimulus did not contribute to switch costs.

Focusing on task-repetition runs only, we measured restart cost as the difference in performance between trials 1 and 2 and mixing cost as the difference in performance on the non-cued trials under mixed-tasks conditions (Experiments 1 and 2) and single-task conditions (Experiment 3). The restart cost was observed under mixed-tasks conditions with both unpredictable and predictable task orders but not under the single-task condition. In contrast, the mixing cost was observed under the mixed-tasks condition with unpredictable task order only (Experiment 1). This finding implies that the optimal task execution on repetition trials depends on how predictable the identity of the approaching task is.