One problem with GUIs from the start has been that automating their behavior is fragile.
Automating actual keystrokes and mouse movements is the most fragile and kludgy of automations. Even drilling to the level of system messages is problematic.
Aside from it's other good qualities, the web is really nice for reducing possible user interactions to a very codified set of operations.
The article here points to ... the wrong way of doing things
That just leaves the other 90% of GUI changes that involve moving a setting to another screen, or changing the way an entire interaction works. ;-)
All the stuff you said about GUI automation being klunky is true. Still, this thing could come in handy as a kind of "GUI batch file," or perhaps as a tool to help produce screencasts.
I don't think any program guarantees that the GUI is a stable API! I do wonder about its ability to handle unexpected conditions, or for the scripter to know that these conditions exist in the first case. For example, if the GUI changes based on the day of week, the script will probably either get very confused, or do the wrong thing. I imagine there's some way to do conditionals (if you see this, then do thing 1, otherwise do thing 2) but it's hard for the scripter to discover all the possibilities. Still, might be useful for some things.