This study investigated how accuracy is attained in fast goal-directed arm movements. Subjects were instructed to make arm extension movements over three different distances in random order, with and without visual feedback. Target width was varied proportionally with distance. Movement time was kept as short as possible, but there were well-defined limits with respect to accuracy. There appeared to be a large relative variability (variation coefficient [VC]) in the initial acceleration. The VC in the distance the hand moved during the acceleration phase was much smaller. This reduction was accompanied by a strong negative correlation between the initial acceleration and the duration of the acceleration phase. Further, the VC in the total distance moved was less than the VC in the distance moved during acceleration. This result indicates asymmetry between the acceleration and the deceleration phase. This is confirmed by the negative correlation between the distance the hand moved during acceleration and the distance it moved during deceleration. Withdrawal of visual feedback had a significant effect on movement accuracy. No differences were found in the parameters of the acceleration phase in the two feedback conditions, however. our results point to the existence of a powerful variability compensating mechanism within the acceleration phase. This mechanism seems to be independent of visual feedback; this suggests that efferent information (efference copies) and/or proprioceptive information is/are responsible for the timing of agonist and antagonist activation. The asymmetry between the acceleration and deceleration phase contributes to a reduction in the relative variability in the total distance moved. The fact that the withdrawal of visual feedback affected movement variability only during the deceleration phase indicates that visual information is used in the adjustment of antagonist activity.