I'm not sure I understand everything but my rough understanding of FRPNow! is ...
Rather than model events always as a stream of events as in most other FRP papers I have read, they consider events individually as a sort of future or promise. The Event acts in many ways like a Behavior of Maybe and they even provide a function "fromJust" which converts a Behavior of Maybe to an Event.
When an async operation starts, the Event is marked as not having occurred yet and having neither time nor value (or infinite future time and undefined value). When the async operation completes, the Event is marked complete and has a timestamp and value.
To create an event stream ala Classic FRP events, they can start with an initial event (say, "read keyboard") and then "unfold" the stream so that the "tail" of the stream is a behavior that depends on the "head" of stream being already available and re-runs the same IO operation over and over.
Behaviors are actually modeled in terms of events - there's the "current" value and an Event for the "future" behavior which will have a new current value and a new Event for the future (recursively).
The rest of the work seems to be how to make this model work efficiently in Haskell without space leaks, which does seem to introduce a fair amount of incidental complexity.
Links: