Instrument every LLM call with traces, token/cost metrics, and latency per turn. Without it, debugging non-deterministic behaviour is guesswork and cost overruns go unnoticed. Consider tools like Langfuse, LangSmith etc.
