Very cool!! I'll give it a spin! I've been using a similar approach for building a tool for agents to drive a running instance of MS word. The closed-testing loops are very powerful with agents.

I had this problem too - I just found that the inconsistency of the the LLM vision + pressure of either having to write OOXML from scratch, or control the app via computer use just didn't really give great results (mostly words fault). Ultimately, for non browser DOCX creation I had more success using some kind of IR that compiles to OOXML.