The 8-15 Hz thalamocortical oscillations known as sleep spindles are a universal feature of mammalian non-REM sleep, during which they are presumed to shape activity-dependent plasticity in neocortical networks. The cortex is hypothesized to contribute to initiation and termination of spindles, but the mechanisms by which it implements these roles are unknown. We used dual-site local field potential and multiple single-unit recordings in the thalamic reticular nucleus (TRN) and medial prefrontal cortex (mPFC) of freely behaving rats at rest to investigate thalamocortical network dynamics during natural sleep spindles. During each spindle epoch, oscillatory activity in mPFC and TRN increased in frequency from onset to offset, accompanied by a consistent phase precession of TRN spike times relative to the cortical oscillation. In mPFC, the firing probability of putative pyramidal cells was highest at spindle initiation and termination times. We thus identified "early" and "late" cell subpopulations and found that they had distinct properties: early cells generally fired in synchrony with TRN spikes, whereas late cells fired in antiphase to TRN activity and also had higher firing rates than early cells. The accelerating and highly structured temporal pattern of thalamocortical network activity over the course of spindles therefore reflects the engagement of distinct subnetworks at specific times across spindle epochs. We propose that early cortical cells serve a synchronizing role in the initiation and propagation of spindle activity, whereas the subsequent recruitment of late cells actively antagonizes the thalamic spindle generator by providing asynchronous feedback.