As you noticed, you must build systems for deterministic work and systems to shuttle that work in an out of an LLM and the result should be no instances of the LLM thrashing about looking for context it’s missing
There’s no single answer other than architecture
-- {no channel}