[stuven]

That's a good point, and I guess the meta point is that there are a thousand and one things I haven't mentioned here that it would have to do as well. To some extent perhaps we could solve this from the LLM the same way a human would: by iteratively plugging in the errors generated by the output back into the bot and taking some next step based on the suggestions generated.

But then we'd have to coerce the bot into generating structured responses that can act as next steps.