The selection of the proper reference class is a serious problem for
anthropic reasoning. But perhaps anthropic reasoning is not necessary to
take advantage of a theory of everything.
Consider how an non-sentient being (excluded by most proposals for the
reference class) can use a TOE. Imagine a non-sentient oracle that was
built to accomplish some goal (for example to maximize some definition of
social welfare) by answering questions people send it. The oracle could
work by first locating all instances of itself with significant measure
which are about the answer the question it's considering (by simulating
all possible worlds and looking for itself in the simulations). Then for
every potential answer, it computes the approximate consequences for the
meta-universe if all of its instances were to give that answer. Finally
it gives the answer that would maximize the value function.
Sentient beings can follow the same decision procedure used by the oracle.
Suppose you are faced with a bet involving a tossed coin. There is no need
to consider probabilistic questions like "what is the probability that the
coin landed heads?" which would involve anthropic reasoning. You know that
there are worlds where it landed heads and worlds where it landed tails,
and that there are instances of you in all of these worlds which are
indistinguishable from each other from the first-person perspective. You
can make the decision by considering the consequences of each choice if
all instances of you were to make that choice.