AMDG

So far I see only that it adds more complexity. 1) You introduced circular dependency between to_ostream operation and arguments 2) The any<> now takes argument wrappers rather than a list of operations

What about this? struct to_ostream : function<to_ostream, any_arg<0>&(any_arg<0>&, const any_arg<1>&)> { //... }; typedef boost::mpl::at_c<to_ostream, 0>::type ostream; typedef boost::mpl::at_c<to_ostream, 1>::type ostreamable; Using to_ostream directly will get both of these. Each any_arg<...> can be used more than once: struct plus : function<plus, any_arg<0>(const any_arg<0>&, const any_arg<0>&)> { //... }; At this point typedef any_arg<0> anyT; yields the original interface.

If you need an operation that takes anys of different types why don't you do it in a straightforward manner?

// Interface for storing reference_wrapper<ostream&> objects typedef any<ostream_ref_operations> ostream_ref;

struct to_ostream : function<ostream_ref&(ostream_ref&, anyT const&)> { // ... };

In this case that works but it doesn't allow the other operations of ostream to vary. Further, it prevents dispatching on the ostream since I don't know that dispatching is needed. The main implementation problem I have is: struct f : function<f, any_arg<0>(const any_arg<0>&, const any_arg<1>&)> {}; template<> struct overload<f, int> { typedef boost::mpl::vector< float(int, long), //ok - we know enough to instantiate all the operations of arg<f, 0> for float char(char, int), //ok - the operations for char are already instantiated bool(float, int) //illegal - can't instantiate all the operations for bool > type; }; at_c<f, 1>::type doesn't know what other operations at_c<f, 0>::type uses. In Christ, Steven Watanabe