You cannot evaluate a voice agent without knowing what it runs on, so this is the full list rather than a diagram with unlabelled boxes. Reasoning, speech recognition, synthesis, telephony, retrieval, observability, and the fallback path each provider hands off to when it degrades.
A single-provider agent is one incident away from a silent phone line. Health checks run per request, and a degraded primary is bypassed mid-conversation rather than after the call fails. The trade-off is that answers can vary slightly in phrasing between providers, which is why the eval suite runs against every model in the chain, not just the primary.
Per-minute cost is dominated by synthesis and reasoning, not telephony. Cartesia sits at roughly half of ElevenLabs per minute at comparable warmth, which is why it is the default. Retrieval and logging are effectively fixed cost. Published per-minute figures from voice API vendors exclude the reasoning layer, which is the single largest line item.