Monkeys were trained to track moving visual targets using a hand-held joystick. The overall frequency response of their visuomotor system was determined using sinusoidal target waveforms. Their responses could be approximately represented by a linear feedforward model consisting of a 0.9 Hz low-pass filter with an additional 150 ms time delay. However, the monkeys normally tracked the target by making intermittent movements of the joystick. Thus, their responses were more realistically modelled as a non-linear sampled feedback model with a loop delay of 250-280 ms. Intermittency allows the monkeys to achieve a good frequency response and maintain tracking stability despite an irreducible visuomotor loop delay of 250-300 ms. When tracking pseudorandom waveforms the monkeys' movements were mainly controlled by positional error. But when tracking predictable sinusoids the amplitude and velocity of each movement was not solely determined by positional error. Instead the monkeys made use of target feedforward, and also internal models of the target waveform, in order to improve their tracking performance. Feedforward control dominated feedback control at high target frequencies, suggesting that the monkeys cannot model targets with long cycle periods.