Generally it's done offline.
Basically, you set up your Maps, give all the information to the A* algorithms, then run a system whereby popular tracks are recorded.
(That's simply done by lowering the cost of cells that have already been visited, and slowly, pattens will emerge)
Once this is complete, you can store that information along with the A* Map info, and load it in at run-time.
Then, using this 'frequent path' information, you can quickly get from one larger cell to the other, whilst still having the dynanic generation of the lower-level (more detailed) cells
'tis a bit of a neuscance(sp?) to run it offline, but for most games, the map doesn't change that often to require it to be dynamic, and most RTS games like StarCraft, etc, simply generate their A* algo's offline as the levels are made.
Also, if your terrain does change, then you can simply run quick checks along the lines of collision with an unexpected object, and then do a very fast path-find around that back onto the main track it was following.
There's lots of theory behind all of this, and I'm sure you, like me, have read most of it whilst developing your A* solution, but I can't reccomend GameDev.net and well-thought-out google searches enough!
Mull it over, and write it out on paper, do up some diagrams, and see if you can work out the systems to use... Pretty simple compared to the A* itself, but still complicated!
Jess.
