Bayesianism works well with known unknowns. But it doesn't work any better than any other system else with unknown unknowns. I would say that while Bayesian reasoning can deal well with risk, it's not great with uncertainty - that's not to say uncertainty invalidates Bayesianism, only to say that Bayesianism is not so spectacularly strong it is able to overwhelm such fundamental difficulties of epistemology.

To my mind, using multiple models of reality is more or less essential. My reasons for thinking this are difficult to articulate because they're mired in deep intuitions of mine I don't understand very well, but an analogy might help somewhat.

Think of the universe's workings as a large and enormously complicated jigsaw puzzle. At least for human beings, when trying to solve a jigsaw puzzle, focusing exclusively on the overall picture and how each individual puzzle piece integrates into it is an inefficient process. You're better off thinking of the puzzle as several separate puzzles instead, and working with clusters of pieces.

By doing this, you'll make mistakes - one of your clusters might actually be upside down or sideways, in a way that won't be consistent with the overall picture's orientation. However, this drawback can be countered as long as you don't look at the puzzle exclusively in terms of the individual clustered pieces. A mixed view is best.

Maybe a sufficiently advanced AI would be able to most efficiently sort through the puzzle of the universe in a more rigid manner. But IMO, what evidence we currently have about intelligence suggests the opposite. AI that's worthy of the name will probably heuristically optimize on multiple levels at once, as that capability's one of the greatest strengths machine-learning has so far offered us.