With sufficiently fast data sampling, ubiquitous sharp transients appear in magnetoencephalography (MEG) data. Initially, no known collective neuronal activity could explain MEG signal generation well above 100 Hz, so it was assumed that these transients were entirely composed of background electronic noise that could be eliminated by filtering and averaging. Recent studies at the cellular level provided evidence for synchronous synaptic input to dendrites and volleys of near-simultaneous action potentials. MEG studies have also identified high-frequency oscillations well above 200 Hz after averaging large number of somatosensory evoked responses. In this study, we searched for evidence of high-frequency neuronal activity in the raw MEG signal using the highest sampling rate available with our hardware. Two human subjects participated in three experiments using visual cues to define planning, preparation, and execution or inhibition of saccades. Tomographic analysis identified "MEG spikes" that were widely distributed across the cortex, cerebellum, and brainstem during cue presentations and saccades. Here we demonstrate how these MEG spikes can be recorded and localized in real time and show that task demands influence their properties. The MEG spikes were organized into feedforward and corollary discharge sequences that could, when combined with the slower activity-linked processing in discrete brain areas over long periods, lasting hundreds of milliseconds. Preparation for impending saccade began as soon as relevant information became available. Cues providing partial information initiated competing motor programs for as yet undecided future actions that were maintained until cues with new information resolved the uncertainty.