realtime-voice-agent-turn-taking-stack

This project presents a realtime voice agent turn-taking stack. It is organized around three code areas: agent/, harness/, and evals/. The page states that "agent/ speaks, harness/ calls it, evals/ judges it. Nothing reaches a caller that the suite has not passed." The agent is described as a router, not a brain: "The agent is a router, not a brain: turn-taking, tool choice, and personality at temperature 0. Real knowledge goes through check_notes , a consult brain measured at 10-22s end to end."

The stack includes a 61-case evaluation suite, text and voice probes, production metrics, and watchdogs. It reports service level objectives for answer latency, tool latency, tool failure rate, clean close, and dead-air. The page says "Two SLOs breached over the lifetime, both fixed , 61/61 green" and that "Lifetime covers all 119 production calls since mid July; 1,007 synthetic harness conversations are excluded so the instrument never grades itself."

The system has one customer on their own number, with accounts kept apart, billing real calls since mid July. It supports capabilities such as sweeping communications, reading back a calendar, scheduling wake-up calls, splitting bills, giving weather and distance, telling jokes, and describing its own capabilities. The page notes "One customer today, on their own number, with accounts kept apart, billing real calls since mid July."