Voice agent passes 15 metrics but customers still complain — what's missing?
Ok-Panda5537 · reddit · 2026-11-15
A developer describes their voice-agent eval setup: 15 metrics including task completion, tool-call success, latency, and several LLM judges — all green — yet support keeps forwarding complaints from calls that look fine by those numbers.
Manually listening to calls revealed real failure modes no metric covered: the agent repeating itself after long pauses, and a confusing way of offering appointment times. The poster asks how others discover failure modes before metrics exist for them — whether people cluster failed calls or audit batches.
More from coding & agent
- Perplexity API adds max_staleness to force fresh crawls at $0.001 per fetch — VeryWellVersed · 2026-11-15
- 'Keep screenshotting until it hits 9/10': the self-iterating AI workflow people are testing — lxfater · 2026-11-15
- Indie dev shows off T3 Code dashboard with 452 settled agent threads — yihui_indie · 2026-11-15
- Amazon planning World Product Graph to index the web's products for AI shopping agents — jomic01 · 2026-11-15
- Google Labs open-sources 15 AI Design Skills that turn prompts into React UI, 9k GitHub stars — tokumin · 2026-11-15
- 20 GitHub repos turn Fable 5.2 into a full motion studio — 3D, FX, sound, export — PrajwalTomar_ · 2026-11-15