I don't recall seeing that implemented in Racket/Scheme, but, in class,
years ago, Leslie Kaelbling mentioned using Scheme captured
continuations for AI search backtracking, as I mentioned (and Matthias
has good comments in that thread):
https://groups.google.com/d/msg/racket-users/jHPtw3Gzqk4/AAqsc-x-AgAJ
That might've been in the class for which we were using a draft of the
Russell&Norvig text; I don't know whether it was mentioned in there.
IIRC, Leslie at the time was coming from Stanford AI Lab tradition (West
Coast tradition OG, to MIT's East Coast).
For most purposes, perhaps one would probably want to write a search two
different ways: one that takes advantage of Scheme's first-class
continuations, and one that doesn't; and compare them (both performance
and ease of implementation).
I don't know how relevant to performance it would be that we almost
never see first-class continuations being leveraged directly in users'
code (outside of the implementation of a Scheme itself), and how that
might have affected priorities in the Scheme implementation. The
implementor of a particular Scheme could say.