Most monads for which you can implement a monad transformer are implemented by wrapping the monad transformer around Identity, but that's simply because you would otherwise be implementing basically the same thing twice. My impression has been that they were originally implemented directly and only transformer-ified later. Certainly, there is no reason that couldn't have been done. Identity doesn't involve any particular voodoo:

    newtype Identity a = Identity { runIdentity :: a }

    instance Monad Identity where
        return a = Identity a
        m >>= k = k (runIdentity m)

Haskell originally didn't include monadic IO - it was introduced in version 1.3, IIRC. IO looked like lazy streams of events, but it was pretty messy and pretty possible to deadlock, from what I hear. Writing modern Haskell, doing IO without any use of the IO type isn't possible, but you can - in principle - avoid ever using the fact that it's a monad by simply calling unsafePerformIO immediately on any IO action. Good luck actually ordering your IO operations, though.