Question says most of it. I created a matrix of descriptors, set the vectors of responses, and input a set number of iterations. Each time I run the function with the same exact inputs, I get the same confusion table, but I get a different training error graph and different variable importance plots.
Boosting, together with bagging, falls into the realm of so cold ensemble models: you randomly draw a sample from the data, fit a model, adjust your predictions, sample once again. Unless your samples are fixed, every time you run the algo you'll get slightly different results.