That's even less tractable a problem than summing over the utility functions of all existing agents, but that's not necessarily a game-changer. There are some other odd features of this idea, though:

• It only seems to work with preference utilitarianism; pleasure/pain utilitarianism would still treat the painless death of an agent with neutral expected utility as neutral. Fair enough; preference utilitarianism seems less broken than conventional utilitarianism anyway.
• Contingent on using preference utilitarianism, certain ways of doing the summing lead to odd features regarding changing cultural values: if future preferences are unbounded in time, a big enough stack of dead ancestors with strong enough preferences could render arbitrary social changes unethical. This could be avoided by summing only over potential lifespan, time-discounting in some way, or using some kind of nonstandard aggregation function that takes new information into account.
• Let's say we're now at a point in time . We can plan for using only the preferences of existing or previous agents; all very intuitive so far. But let's say we consider a time further in the future. New agents will have been introduced between and , and there's no obvious way to take their preferences into account; every option gives us potential inconsistencies between optimal actions planned at and optimal actions taken at time . The least bad option seems to be doing a probability-weighted average over agents extant in all possible futures, but (besides being just ridiculously intractable) that seems to introduce some weird acausal effects that I'm not sure I want to deal with. Taking the average at least avoids some of the crazier possible consequences, like the utilitarian go forth and multiply that I'm sure you've thought of already.