Usually when stacking, one would want to try different stacking algorithms (i.e. GLM or Neural Net) and then select the best one. Currently in automl, only one stacked model is trained, thus requiring users to manually train other stacked models. It would be better to abstract this into the API.
Multi-level stacking is the idea that after fitting > 1 stacked model, we stack the stacked model. This may not improve metrics by much, but could be useful.