In this article, we performed an analysis of the dependability of an elementary yet critical robot component, i.e., the joint-level actuation subsystem. We consider robot actuators that implement the VSA paradigm, i.e., ability to change the effective transmission stiffness during motion to achieve high performance while constantly keeping injury risks by accidental impacts with humans below a given threshold. Without attempting a comprehensive review of different existing design approaches to VSA,we focused on the analysis of three different arrangements of agonistic/antagonistic actuation mechanisms for pHRI applications. Several aspects of their performance, safety, and dependability have been considered to get an indicative, though certainly not exhaustive, comparison of these alternatives. According to our results, the simple AA arrangement is more reliable (due to the simplicity of its mechanical implementation) if FM is not used. Proper FM actions can make other designs perform equally well as the simple AA concerning reliability and can perform better for steerability. Simulations of impacts in failed states (where FMis not used by a worst-case assumption) also show that the different designs have comparable safety properties. Although overall results for the bidirectional arrangements are somewhat superior, especially in terms of steerability (if FM is applied), we do not extrapolate any general claim in this regard. Indeed, many factors influence the results of similar studies, and each case should be considered in detail and very carefully. The scope of the study can become quite broad, and many of the theoretical and technical issues presented here (e.g., fault detection, supervisory control, and safety-related systems) will require further separated investigations. One of the purposes of this work was to explore and further promote dependability studies in robotics, as a means of addressing concerns in safety-critical robotic systems for physical intera- - ctions with humans. In this sense, a robot for pHRI applications is a unique benchmark for improving the state of art of fault tolerant design as well as in developing tools to master performance, dependability, and safety issues of a robotic structure.