Shipped a voice agent on the Realtime API. Went through production call logs and found 7 behavioral bugs that no amount of scripted testing would have caught

A developer deployed a voice agent using the Realtime API and, after reviewing production call logs, identified 7 behavioral bugs that scripted testing failed to catch. One significant bug involved LangGraph's fan-out mechanism, where multiple distinct requests in a single turn led to parallel writes. This exposed a race condition where response fragments from parallel branches could return out of order, violating formatting rules. The fix involved tagging messages by their originating node and sorting them deterministically, a solution confirmed against live API calls.

This report uniquely details 7 behavioral bugs found in a voice agent through live call logs, unlike typical scripted testing, including a previously undocumented LangGraph fan-out ordering issue.