The investigation of the brain connectome with functional magnetic resonance imaging (fMRI) and graph theory analyses has recently gained much popularity, but little is known about the robustness of these properties, in particular those derived from active fMRI tasks. Here, we studied the test-retest reliability of brain graphs calculated from 26 healthy participants with three established fMRI experiments (n-back working memory, emotional face-matching, resting state) and two parcellation schemes for node definition (AAL atlas, functional atlas proposed by Power et al.). We compared the intra-class correlation coefficients (ICCs) of five different data processing strategies and demonstrated a superior reliability of task-regression methods with condition-specific regressors. The between-task comparison revealed significantly higher ICCs for resting state relative to the active tasks, and a superiority of the n-back task relative to the face-matching task for global and local network properties. While the mean ICCs were typically lower for the active tasks, overall fair to good reliabilities were detected for global and local connectivity properties, and for the n-back task with both atlases, smallworldness. For all three tasks and atlases, low mean ICCs were seen for the local network properties. However, node-specific good reliabilities were detected for node degree in regions known to be critical for the challenged functions (resting-state: default-mode network nodes, n-back: fronto-parietal nodes, face-matching: limbic nodes). Between-atlas comparison demonstrated significantly higher reliabilities for the functional parcellations for global and local network properties. Our findings can inform the choice of processing strategies, brain atlases and outcome properties for fMRI studies using active tasks, graph theory methods, and within-subject designs, in particular future pharmaco-fMRI studies.