Monkeys were trained to make saccades to briefly flashed targets. We presented the flash during smooth pursuit of another target, so that there was a smooth change in eye position after the flash. We could then determine whether the flash-evoked saccades compensated for the intervening smooth eye movements to point the eyes at the position of the flash in space. We defined the "retinal error" as the vector from the position of the eye at the time of the flash to the position of the target. We defined "spatial error" as the vector from the position of the eye at the time of the saccade to the position of the flashed target in space. The direction of the saccade (in polar coordinates) was more highly correlated with the direction of the retinal error than with the direction of the spatial error. Saccade amplitude was also better correlated with the amplitude of the retinal error. We obtained the same results whether the flash was presented during pursuit with the head fixed or during pursuit with combined eye-head movements. Statistical analysis demonstrated that the direction of the saccade was determined only by the retinal error in two of the three monkeys. In the third monkey saccade direction was determined primarily by retinal error but had a consistent bias toward spatial error. The bias can be attributed to this monkey's earlier practice in which the flashed target was reilluminated so he could ultimately make a saccade to the correct position in space. These data suggest that the saccade generator does not normally use nonvisual feedback about smooth changes in eye or gaze position. In two monkeys we also provided sequential target flashes during pursuit with the second flash timed so that it occurred just before the first saccade. As above, the first saccade was appropriate for the retinal error provided by the first flash. The second saccade compensated for the first and pointed the eyes at the position of the second target in space. We conclude, as others have before (12, 21), that the saccade generator receives feedback about its own output, saccades. Our results require revision of existing models of the neural network that generates saccades. We suggest two models that retain the use of internal feedback suggested by others. We favor a model that accounts for our data by assuming that internal feedback originates directly from the output of the saccade generator and reports only saccadic changes in eye position.