Building a phone voice agent: telephony and barge-in ate the time, not the LLM

Fish0plays · reddit · 2026-09-06

A postmortem on a team's first production phone-based AI agent: only 15% of engineering time went to conversation behavior; the rest went to unglamorous infrastructure — SIP setup, call routing, and edge cases; reliable turn detection and barge-in on real phone calls; observability (dumping transcripts to raw JSON logs was useless for debugging); and failure handling (timeouts, dropped calls, slow tools, weird STT output). Halfway through they evaluated managed platforms including Vapi, Retell, and Dasha, and in retrospect would have started with a managed runtime to focus engineering time on business logic.

Original post →

More from coding & agent

coding & agent channel →