Voice Agents Fail at ~10%, and Adversarial Tests Break 1 in 5

AI Engineer · youtube · 2026-09-15

Hamming AI founder Sumanyu Sharma argues voice agents are dangerous not because of any single failure but because they are centralized and scaling fast. He was told by a voice agent that a doctor's appointment was booked; he wasn't on the schedule and lost two hours. Make that a grandparent and a procedure instead of a checkup, and the same failure costs far more.

His framing: crime is decreasing and local, touching only those involved; voice agents are centralized, so one prompt change propagates to everyone at once. Around a trillion phone calls happen yearly — even a 1% error rate means 10 billion bad interactions, and across the 10,000 agents his company monitors the real rate is nearer 10%.

Failures are rarely dramatic: an agent reports it found the right policy while quietly skipping the eligibility check, applies an unauthorized discount, or claims it booked something it did not.

His remedy is a loop, not a fix: find problems, size them by frequency and severity, change something, verify the change did not break anything else, and keep watching. Listening to individual calls by hand is where to start, not what to scale — the real insight lives in patterns across conversations. His team's adversarial testing breaks roughly one agent in five.

Original post →

More from coding & agent

coding & agent channel →