→ Latency under load. A voice agent that feels instant in a demo can fall apart at real concurrency. Design for tail latency, not average latency.
→ Reliability and guardrails. Hallucinated responses are an annoyance in chat and a liability on a phone call. Production systems need guardrails, deterministic fallbacks, and a clean path to human handoff baked in from day one.
→ Observability. Treat monitoring, evals, and call-level visibility as launch requirements, not post-incident cleanup.
Cheers to our hosts for pulling the room together, and see you at the next one!