I've seen this exact issue before it's usually the LLM temperature being too high, so the agent gives slightly different responses each time instead of staying consistent. Quick question though are you using a custom LLM or one of Retell's built-in models?"