Sumanyu Sharma argues that a voice agent sounding human is not evidence that it performs its task correctly. Calls can sound fluent while supplying wrong information or confirming actions that never happened. Sumanyu Sharma attributes an approximately 10% error rate to the agents monitored by Hamming AI, rather than presenting that sample as an industry-wide measurement.
Sumanyu Sharma proposes finding failures across conversations, prioritising them by frequency and severity, diagnosing their causes, and testing fixes for regressions. Centralised agent prompts and shared tools make a single change capable of affecting many calls, so realistic variations in wording, accents and intent matter more than replaying one exact script.
Sumanyu Sharma distinguishes ordinary reliability testing from adversarial testing of agents that can access tools and sensitive data. The talk recommends testing before deployment, comparing changes against real outcomes, and continuously monitoring production calls and high-risk behaviours.
Watch on YouTube
