video1.0<iframe src="https://www.loom.com/embed/f81fc1f273c24ccaaf26f62766cb6b9e" frameborder="0" width="1920" height="1440" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>14401920Loomhttps://www.loom.com14401920https://cdn.loom.com/sessions/thumbnails/f81fc1f273c24ccaaf26f62766cb6b9e-6e3bf28ea7212944.gif577.4This Loom demonstrates a policy-driven refund agent with live voice integration. The presenter shows a happy path refund approval for order ORD 1001 and an edge case denial for KC’s order 1003 due to final sale under policy R2, plus a decline when the order does not belong to the requesting customer under R6 and a refusal to handle non-refund requests. The system runs the same six tool calls and deterministic policy engine for both text chat and voice using OpenAI’s real-time API over WebRTC, with tool calls executed server-side via a single executor. It includes failure handling with retries, a simulated 503 flow, and escalation to a human per policy R4, and notes adversarial sweep loops with 44 findings logged in sweeps.md.