There's a long history of ``clever new ways'' to use existing crypto algorithms turning out to have serious flaws -- a good example is early attempts to improve the strength of (56-bit key) DES by encrypting three times with three keys. This turns out to introduce enough non-randomness to make the result much weaker than one might expect; standard 3DES works by encrypting with the first key, decrypting with the second, and encrypting with the third, which results in very different properties of the output cyphertext.
I'm not saying BCrypt has the same sort of issues, but I'd like to see some cryptanalysis of this before trusting my users' data with it. Notably, there seems to be no links to such analysis on the BCrypt home page -- not even an argument from the author as to why this code should be cryptographically sound.