MTVA-Bench finds voice agents miss the script

MTVA-Bench evaluated 49 voice agents across 490 reviewed scenarios and found its seven-model study’s overall scores spread 24.4 points even though correct-tool selection stayed within 6.4 points, per the arXiv paper. The benchmark weights task completion and conversation quality equally, checking arguments, action order, script compliance, and what agents say around tool calls. Teams building on Gemini Live’s history-cost model should file this as a reminder: passing tool-choice tests is not the same as running a good call.