So he created the Gosling project. It's pretty embryonic - and I think some of the file and resource classes could be replaced by straight usage of things like javax.tools.FileObject, but it has a nice simplicity. Here's the constructor for Gosling's own build to build itself. It has a similar feel to what Jon did in designing Wicket's APIs:

    public GoslingProject() {
        final Folder root = new Folder("/Projects/gosling/workspace/gosling");
        final Folder source = root.folder("src");
        final Folder lib = root.folder("lib");
        final Folder target = root.folder("target");
        add(new JavaApplicationBuilder() {
            @Override
            protected Set getDependencies() {
                final ResourceDescriptorSet dependencies = new ResourceDescriptorSet();
                dependencies.add(Apache.apache.wicket.core.development.resources());
                dependencies.add(Apache.apache.wicket.extensions.development.resources());
                return dependencies;
            }

            @Override
            protected Set getJars() {
                return lib.nestedJars();
            }

            @Override
            protected Folder getSourceFolder() {
                return source;
            }

            @Override
            protected Folder getTargetFolder() {
                return target;
            }
        });
}

I mean, the appeal of things like Ant is that many things are built in a pretty similar way; Maven is even more "my way or the highway" in that regard. I use (and sometimes like) both of them. And in theory an Ant XML script (xml script - now there's an oxymoron) is human readable - although I challenge anyone to make heads or tails of this in ten minutes. It certainly beats Make and tab vs. space madness. There's an argument that it's toolable - that a tool can analyze an Ant script. This seems to me to be a red-herring if you do a design like Wicket or Gosling use - where you know that the entire state you need to analyze is going to be set up in the constructor of a known class of a known type. The Javac Tree API may not be for the faint of heart, but analyzing the closure of a constructor is perfectly doable.

My point is that Ant doesn't really deliver the clarity it promised except in the most trivial of cases. You could have at least as much clarity with plain-old Java code - you just need to start from a good design so the code can speak for itself. And a design where all targets will be added to the build in the constructor is pretty darned clear. Isn't one of the things the agile crowd talks about a lot letting code speak for itself?

Not to mention that running javac in-process ought to be blazingly fast. And that such a project could import and call existing Ant tasks to do things with a thin adapter layer - so anybody's custom tasks or missing functionality could be handled leveraging stuff that's already out there.

It seems like a pretty nice idea to me.