§2.3 · Research thread · active · proprietary
Split-second decisions
Kansas City · Independent ·
Abstract
The small decisions a voice agent makes in every turn of a call, answered by something small and fast. The result is public in relative terms: how often the small decisions were right, how fast, and how far the whole call still is from natural. The measured figures are in the full results paper, shared on request. Which models were used and how the questions are worded are not published. It is one untuned run on an easy, machine-made test in one setting; the end-to-end timing is from a scripted call.
1 · Article2 · PaperRequest access
Keywords: voice, agents, latency.
1Measurements
| measure | value | conditions | source |
|---|---|---|---|
| Decisions right | 39 in 40 | about; mean across the kinds of decision, first run, before any tuning | private research log: voice agent lab |
| Time per decision | < 1/4 | of the time allowed, typical case | private research log: voice agent lab |
| Weakest kind of decision | 9 in 10 | about; checking a drafted sentence against a policy | private research log: voice agent lab |
| Caller's wait, end to end | ≈ 3× | the target for natural conversation; the decisions are about a tenth of it | private research log: voice agent lab |
§Where the work lives
§Related writing and lessons
- Thinking in Agents: The New Mental Model Every Developer Needs (writing)
- UI Out of Thin Air — What Frontend Engineers Will Actually Do After Gemma 4 (writing)
Ask this thread