Updated November 24, 2009
The current approach is to get up through the first level analyses in FSL using the nipype system. If you are a bad programmer, you might have a look at Satra's FSL tutorial (note, as of this writing, the newer docs have yet to make their way there - so have a look at the SPM tutorial if FSL isn't covered there). If you are a mediocre or better programmer, or otherwise would like to see one of my actual pipelines, please contact me!
After getting the first-level analyses done, it is straightforward to use the Feat GUI to run a set of contrasts across an arbitrary number of first-level fits, and this process is relatively fast. Also, if you load up new first-level directories (or even load from an existing design.fsf), the Feat GUI will remember the contrasts you entered for the previous batch.
Higher level analyses are relatively straightforward to do from there - I do a single group average (using fixed effects) across the scans for a given subject. Then, I do a FLAME 1 average across each contrast separately.
Unlike running contrasts, the Feat GUI will not batch higher-level analyses. So, I usually save all of the analyses I'd like to do to disk, and then run a semicolon separated set of commands like so:
feat subj1.fsf; feat subj2.fsf; etc.
You probably should have a little over 3GB and a free CPU core per first-level analysis that will be run in parallel. In some cases, your processes may compete for disk IO as well, which could slow things down, so that's another thing to look at. Unfortunately, when FSL runs out of memory, it just dies. Fortunately, if you are using nipype, you will pick up right where you left off - so you shouldn't lose too much.