TV noise crashes voice agents: task completion falls from 38.7% to 8.6% across 14 stacks
solyarisoftware · x · 2026-10-14
joshmeyer benchmarked 14 voice agent stacks with a TV playing in the background: overall task completion fell from 38.7% on a clean line to 8.6%. Per-stack results vary widely: gpt-live 1 holds up best (59.1% → 36.1%), grok voice stays strong (39.7% → 33.8%), while deepgram, vapi and livekit drop to 0%, and openai realtime variants crash from 49% to 3-4%. Based on 961 clean and 945 TV-condition calls. A stark reminder that real-world noise can gut voice agent reliability.
More from coding & agent
- Anthropic Post: Claude Rewrote Pipeline With 20x Less Memory and 4x Speed — teortaxesTex · 2026-10-14
- Cursor now ships real Supabase backend changes via MCP, no SQL needed — XFreeze · 2026-10-14
- Developer: superhuman agents finished the last 20%, so vibe coding is dead — josh_bickett · 2026-10-14
- MIRA-AC trains research agents on proxy feedback, lifting symbolic regression from 50 to 63 solved — anirudhg9119 · 2026-10-14
- MIRA-AC behavior shift: hypothesis-distinguishing tests up 16.8% to 28%, stalled tweaks halved — anirudhg9119 · 2026-10-14
- MIRA-AC RL details: stitching decisions into one logical episode, updating only the meta-reasoner — anirudhg9119 · 2026-10-14