1. Rhesus monkeys were trained to look to brief visual targets presented in an otherwise darkened room. On some trials, after the visual target was extinguished but before a saccade to it could be initiated, the eyes were driven to another orbital position by microstimulation of the paramedian pontine reticular formation. If, as current models of the saccadic system suggest, a copy of the motor command is used as a feedback signal of eye position, failure to compensate for stimulation-induced movements would indicate that stimulation occurred at a site beyond the point from which the eye position signal was derived. 2. Animals compensated for perturbations of eye position induced by stimulation of most pontine sites by making saccades that directed gaze to the position of the visual target. With stimulation at other pontine sites, compensatory saccades did not occur. 3. Pontine stimulation sometimes triggered, prematurely, impending visually directed saccades. The direction and amplitude of the premature movement depended upon the location of the briefly presented visual target. The amplitude of the premature movement was also a function of the interval between the stimulation train and the impending saccade. These data suggest that input signals for the horizontal and vertical pulse/step generators develop gradually during the presaccadic interval. Saccade trigger signals need to be delayed until the formation of these signals is completed. 4. The implications of these findings for models of the saccadic system are discussed. Robinson's local feedback model of the saccadic system can explain compensation for pontine stimulation-induced changes in eye position but cannot easily account for the failure to compensate for perturbations in eye position produced by stimulation at other sites. Modified versions of Robinson's model, which assume that the input signal to the pulse/step generator is the desired displacement of the eye, can account for both compensation and the failure to compensate since two separate neural integrators are employed. However, these models ignore kinematic arguments that commands to the extraocular muscles must specify the absolute position of the eye in the orbit rather than a relative movement from a previous position.