The author of numpy, Travis Oliphant recently moved from Utah to Texas to work with Enthought. One of the many contributions Enthought has made to the scientific software community is developing a whole suite of tools (in Python) that make developing cross-platform scientific applications much easier. In a recenet post, Travis talks at length about what makes a "good" UI. Travis has obviously thought a lot about UI's for scientific computing and his discussion is very interesting. Definitely worth a read. Briefly, Travis notes that the user interface of an app is not just the buttons and menus, but the entire application framework—persistence, undo/redo, safe exploration, workflow, etc.
In scientific software, I think "workflow" is the most important part of an application's user interface. Application frameworks such as Apple's Cocoa, Microsoft's .NET, Trolltech's Qt, etc. have solved many of the applicaiton framework UI issues that Travis mentions (in fact, this is why we use Cocoa and Qt for most of our work at Physion). Apple's Cocoa framework is particularly impressive in this regard. An application with undo/redo, persistence to an SQLite database, automatic network discovery of distributed computing resources, you-name-it, is virtually code-free for the developer. But, none of these frameworks have solved the scientific workflow problem. That's because we know how undo/redo should work. We've known for decades how a database should work (for the most part), but science is about the new and very often discovering something new involves creating the entire workflow anew.
One reason why new workflow and new experiments often go together is because a workflow implicitly defines a world model—an idea about what objects in the world mater and how they interact—and new experiments also define a new world model. Therefore, new experiment, new workflow. Only by matching the workflow to the experiment, can we make software that becomes invisible to the researcher while facilitating new discoveries.