What with this whole test-first thing, some people's attention has drifted away from capture/replay GUI testing tools. I think that's good. But I was reminded today of an oldish article from Mark Silverstein on doing capture below the GUI (pdf). It relies on a clear separation between GUI and guts, with the GUI talking to the guts using the Command pattern.

If you're going to do capture, that seems the way to do it, especially since it's yet another way to encourage layering. If - as is extremely likely - the resulting scripts will be read by humans, not just executed, it might also encourage good (business-facing) names for the guts API.