From last one year, I'm working on a modeling tool featuring support for UML2 and BPMN1.0 specifications. Well, this tool is not envisioned just like "yet another modeling tool" allowing you to draw basic elements (and strange visual manipulations) and generate structural code and documentation (which is not impressively useful!).
It's commonly known that modeling is good for visualizing and documenting a system. Well, I've heavily used UML for documenting and I found it useful for communicating software blueprints to those who don't understand the technology well.
Contemporary tools can do better apart from visualizing, Typically a modeling tool doesn't sponsor full blown generation of an application for target platform, say J2EE; Not that it's insane attempt but it's simply too much of work. They generate structural code (skeletal) from model (say, classes from Class Diagrams) ; this is commonly known as forward engineering. Tools also support round trip engineering for regenerating the code from model while maintaining the manual customizations.
How redundant, Why to model and code at the same time and why maintain both?
The tool I'm working on is supposed to generate "typical application" end-to-end (heh, really?!). Also, Tool has a special language(sort of BASIC dialect) , that can be used to write business logic for the application which might, otherwise, result in tedious micro modeling (boy, you will require incredible mouse capabilities for that level of modeling). This language is customized implementation of OMG(tm) Action Semantics Language (ASL) and Object Constraint Language (OCL); The logic in this language is translated in to platform specific code during transformation and code generation.
Technically, this is enough for generating entire application. Recently, tool vendors are trying to avoid the ASL approach altogether and use supplemental UML models (like activity and state machines) to specify behavioral aspect of system; but this approach has usability problems, although it is an interesting technical challenge (umm... given enough beer and I'll even admit that I wrote a state machine compiler for money to generate behavioral code..).
How bad again, Why I want to learn UML, a dumb ASL language and still struggle with generated code to customize it?
I agree that the tool can generate "student projects" like applications, but still keeps me skeptical about *serious* applications which involve complexities of security, performance and integration. It's not impossible to do that either, however, supporting multiple security/persistence/application frameworks/third party interfaces for multiple platforms, in its entirety, would be a non-trivially huge attempt ever made.
Lately, I'm observing generative methods in regular development; these days, most applications are combination of generated + hand written code(in some reasonable combination), people manage both code as well as model. Generative Technologies are gaining acceptance gradually, especially because support from Eclipse projects like EMF, GEF, GMF and other Modeling subprojects apart from rising community interest, hundreds of modeling tools are available and major players are investing in this technologies.
We have witnessed the transitions in technologies; Transitions from Hardware specific Assembly code to C code, C code to C++/Java/RoR code.., these transisions are fine and gradual, but with modeling being a paradigm shift in the way we see software development (its not just about learning new language, right?), modeling imposes the steep learning curve, and understanding the mammoth specifications. It's always difficult to adopt a different way of doing things.
With the current state of tooling and specifications, it is difficult to offer "Generate everything, Write Nothing" jargon feature, and its pain in a**e to maintain both model and code in sync (tools are improving though). I'm sure of one thing, generative development is going to be the next generation of software development, not sure how many more years...