To this day many good models don't support structured outputs (say Opus 4.5) so it's not a panacea you can count on in production.

The bigger problem is that LangChain/Python is the least set up to take advantage of strong schemas even when you do have it.

Agree about pillaging for prompts though.