One of the most exciting translational prospects for brain imaging research is the potential use of functional magnetic resonance imaging (fMRI) 'biomarkers' to predict an individual's risk of developing a neuropsychiatric disorder or the likelihood of responding to a particular intervention. This proposal depends critically on reliable measurements at the level of the individual. Several previous studies have reported relatively poor reliability of amygdala activation during emotional face processing, a key putative fMRI 'biomarker'. However, the reliability of amygdala connectivity measures is much less well understood. Here, we assessed the reliability of task-modulated coupling between three seed regions (left and right amygdala and the subgenual anterior cingulate cortex) and the dorsomedial frontal/cingulate cortex (DMFC), measured using a psychophysiological interaction analysis in 29 healthy individuals scanned approximately two weeks apart. We performed two runs on each day of three different emotional face-processing tasks: emotion identification, emotion matching, and gender classification. We tested both between-day reliability and within-day (between-run) reliability. We found good-to-excellent within-subject reliability of amygdala-DMFC coupling, both between days (in two tasks), and within day (in one task). This suggests that disorder-relevant regional coupling may be sufficiently reliable to be used as a predictor of treatment response or clinical risk in future clinical studies.