The job demands-resources model (JD-R model; Bakker & Demerouti, 2014) is well established in occupational research, and the proposed processes it posits have been replicated numerous times. Thus, the JD-R model provides an excellent framework for explaining the occupational well-being of beginning teachers—an occupation associated with particularly high levels of strain and consequently, high attrition rates. However, the model’s assumptions have to date mostly been tested piecewise, and seldom on the basis of longitudinal models. With a series of longitudinal autoregressive SEM models (N = 1,700) we tested all assumptions of the JD-R model simultaneously in one model with an applied focus on beginning teachers. We assessed self-reports of beginning teachers at three time waves: at the beginning and end (one and a half to two years later) of their preservice period, and again, one year later. Results revealed significant direct effects of resources (self-efficacy) on engagement, of demands (classroom disturbances) on strain (emotional exhaustion), and a significant reverse path of engagement on self-efficacy. Additionally, the results showed two moderation effects: Self-efficacy buffered the demands-strain relationship, while self-efficacy also predicted engagement, especially when disturbances were high. Thus, self-efficacy in classroom management plays an important role in the teachers’ stress development process, as it will, in case of high classroom disturbances, not only buffer the strain-enhancing effects, but also boost engagement. Commitment was predicted directly by emotional exhaustion and engagement, but indirectly only by self-efficacy (via engagement). Thus, we provide strong empirical support for the JD-R model.