I’m asking specifically about LLM calls embedded inside real production workflows, not demos, side projects, or exploratory prompt work.

Think backend pipelines like: step 1 → LLM → step 2 → LLM → step 3 where users depend on the output and nothing technically “crashes.”

We’ve seen a recurring pattern: - Same input, same prompt, same model - Works reliably for weeks - Then a constraint is ignored, or a later step contradicts an earlier one - Retries don’t reliably fix it - Logs don’t explain what changed

The hardest part isn’t bad output, it’s not being able to explain failures to PMs or stakeholders when nothing obviously broke.

Curious how others operating LLM-backed workflows in production are diagnosing or containing this kind of behavior over time.

(Not looking for prompt advice or eval frameworks. Interested in operational experiences.)