Yeah, this is actually a very common issue with voice agents you’re not doing anything wrong. The inconsistency usually comes from a mix of things: variability in speech-to-text (accents, noise, phrasing), non-deterministic LLM behavior, and silent model updates that slightly change how responses are generated. Without controlled inputs, it’s almost impossible to reproduce bugs reliably. I can help you fix this by making your setup more testable and consistent (things like logging raw audio + transcripts, forcing deterministic settings, and creating replayable test cases). Quick question: are you currently storing the raw audio/transcripts and model parameters for each call, or just the final output?