I don't think you're doing anything wrong. I think we are trying to apply this technology where it doesn't really belong. The whole purpose of something like a

I don't think you're doing anything wrong. I think we are trying to apply this technology where it doesn't really belong. The whole purpose of something like a parser is to reliably capture an AST representation of the thing in question. Asking a statistical model to do the same thing seems insane in principle. You've now turned a deterministic outcome into something that will definitely go wrong with some probability. The fact that LLMs make for terrible parsers is catastrophic for things like function calling. Every successful agentic demo I've seen has had so many guard rails in terms of error feedback and retry that you wonder where the value-add actually resides. As you say, "I'll just convert..."