I understand how COM can achieve compiler agnostic C++ code, since it defines an ABI by being careful what features of the C++ language to use. It's just C++ code talking to C++ code in a really clever way. However I still don't understand how it can allow for language interop with C# or Javascript for example.
Where is the boundary? The only explanation I have right now is that the language compiler itself must have special support for COM so that it can generate the proper assembly code to allow for accurate communication between caller/callees.
Since you've tagged your question with WinRT, I assume you're asking specifically about how this is achieved by WinRT language projections. In that case, all languages must have some way to map their natural language constructs to the COM ABI that WinRT defines. That ABI is derived from metadata encoded in the ECMA 335 standard and special rules are applied to transform the abstract metadata into a concrete ABI. There are naturally different ways to achieve this. The CLR itself was updated to support WinRT in C#. The Visual C++ compiler was (sadly) updated with language extensions to support WinRT via C++/CX. The C++/WinRT approach is very different in that it requires only a standard C++ compiler and all of the knowledge about WinRT is delivered via a standard C++ header-only library. Other languages might take different approaches, but at the end of the day they must agree on the way that types expressed in metadata are transformed into objects and virtual function calls on the ABI based on COM.
And while this process is not well documented at the moment, C++/WinRT is one of the only open source language projections and thus acts as a useful reference implementation for those who need to understand how WinRT works under the hood.