Hey Mieszko. Recently I had a situation where I needed to have path destination updated to reflect a vector in my blackboard, rather than the location of an actual actor (I had to calculate the destination point using an algorithm, from various parameters which included actor locations, and these parameters all changed with time). There are some workarounds of course, like using a proxy actor, or creating a new move request - though this seems to lead to the character stuttering as the new request takes over.
Anyway, it seems that the path observing based on an actor is built deep into the navigation system. It would be nice if this could be abstracted out to the level of observing a path based on something more generic - a delegate which provides the desired destination at a given time, for example. Not sure how big a job this would be though.