A perfect Bayesian agent is really quite simple. It has a database of probability assignments, a utility function, an input system and an output system. Inputs change the probability assignments according to simple rules. It computes which outputs maximize expected utility (either causally or evidentially--it won't matter for this post). And it does that (in cases of ties, it can take the lexicographically first option).
In particular, there is no need for consciousness, freedom, reflection, moral constraints, etc. Moreover, apart perhaps from gerrymandered cases (Newcomb?), for maximizing expected utility of a fixed utility function, the perfect Bayesian agent is as good as one can hope to get.
So, if we are the product of entirely unguided evolution, why did we get consciousness and these other things that the perfect Bayesian agent doesn't need, rather than just a database of probability assignments, a utility function that maximizes reproductive potential, and finely-honed input and output systems? Perhaps it is as some sort of compensation for us not being perfect Bayesian agents. There is an interesting research program available here: find out how these things we have compensate for the shortfalls, say, by allowing lossy compression of the database of probability assignments or providing heuristics in lieu of full optimizations. I think that some of the things the perfect Bayesian agent doesn't need can fit into these categories (some reflection and some moral constraints). But I doubt consciousness is on that list.
Consciousness, I think, points towards a very different utility function than one we would expect in an unguidedly evolutionarily produced system. Say, a utility function where contemplation is a highest good, and our everyday consciousness (and even that of animals) is a mirror of that contemplation.