In this study, we evaluate the use of models simpler than CGCMs to predict decadal climate variations. We have carried out ensembles of century-long climate predictions with a global atmospheric general circulation model coupled to a slab ocean model (AGCM-SOM), with simple extrapolative prescriptions of future greenhouse gas concentrations and other external forcings. The use of the slab ocean model mitigates the problem of climate drift and reduces the impact of initial condition errors in the ocean. In addition to the control integration using the standard AGCM-SOM configuration, we have carried out additional sensitivity experiments where we have eliminated the Wind-Evaporation-SST (WES) feedback, and also allowed the surface heat flux correction (QFLUX), which acts a proxy for oceanic heat transport, to vary over time. Additionally, we have constructed a simple statistical model (a variant of the Hasselmann model) to assess the skill of empirical approaches to decadal climate prediction. Our results suggest that (i) the AGCM-SOM configuration can capture a significant portion of the predictive skill on decadal time scales; (ii) the WES feedback serves to decrease the signal-to-noise ratio over many regions, and (iii) the choice of QFLUX can affect the decadal prediction skill.