From Self-driving to Autonomous Voice Agents — Brooke Hopkins, Coval
Jul 31, 2025 · 19:32
Brooke Hopkins, founder of Coval and former Waymo eval lead, argues that reliability lessons from self-driving cars directly apply to building trustworthy voice agents. She explains how Waymo moved from brittle scenario-specific tests to large-scale probabilistic simulations that measure success rates across thousands of runs—a framework now used at Coval to evaluate conversational AI at scale. Key practices include denoising: resimulating a failed scenario 100 times to assess true failure probability, and calibrating LLM-as-judge metrics with human feedback until automated evals align. Hopkins outlines a hierarchical eval strategy: start with text-based task evals for speed, then graduate to voice-to-voice tests for latency and interruptions, and finally hyper-realistic simulations for production edge cases. She predicts that within three years every enterprise will need a voice experience comparable to a mobile app, making integrated evaluation infrastructure a competitive necessity.
