Call Your Own AI Receptionist and Try to Break It

Google shipped live voice-agent evaluation because a good demo is not evidence. Fifteen minutes of hostile calls tells you what your agent does when the caller ignores the script.

Time
15 min
Difficulty
easy
Potential value
very-high value

1 sources, 1 primary. Verified Sep 1, 2026. Confidence 0.85 of 1.00, where 0.70 is the minimum that publishes.

Why today

Google just shipped native live evaluation for voice agents, on the grounds that getting one into production takes more than a good demo. Most operators running a phone agent have never tested it past the demo, and the failures Google names are the silent kind: tools that stop firing, context that slips between turns, interjections that go ignored. None of those announce themselves. The caller just hangs up.

What you will do

Produce a written list of the exact turns where your voice agent fails, so you are fixing named failures instead of guessing at them.

  1. 01 Call your agent from a number it has never seen, and throw away the demo script. You are testing the conversations real callers have, not the one you rehearsed.
  2. 02 Interrupt it mid-sentence with your actual question. Ignored interjections are a named failure mode, and you will hear within two seconds whether it talks over you or yields.
  3. 03 Change your mind two turns later. Give it a date or a name, let it move on, then correct yourself. This is where context slips between turns, and it is the failure a caller notices most.
  4. 04 Make it actually do something: book, look something up, take a message, transfer. Tools that stop firing are the failure that costs money, and they never announce themselves.
  5. 05 Ask your single most common customer question three different ways, on three separate calls. A spoken reply can be correct in hundreds of phrasings, and so can the question it has to recognise.
  6. 06 Write down the exact turn each failure happened and your exact wording. That list is the deliverable, and it is what you hand to whoever maintains the agent.

How you will know it worked

You have a written list naming at least one turn where the agent failed, with your exact wording, and for each one you can say whether a tool did not fire, context was lost, or an interruption was ignored.

How this was checked

A person wrote this page by hand. Each claim below was checked against the fetched document behind its sources before it published.

Claims checked

  • Google's Agent Development Kit now provides native live evaluation for voice agents.

  • Named failure modes for live voice agents are tools that stop firing, context that slips between turns, and interjections that go ignored.

  • For a live voice agent, timing and recovery matter as much as the content of its answers.

  • Voice agent behaviour that sounded correct yesterday can change quietly on the next prompt tweak or model iteration.

  • Putting a live agent into production requires repeatable evidence that it holds up across the conversations real users will actually have, rather than a good demo.

Research run manual-2026-09-01-voice-agent-adversarial-test. Verified Sep 1, 2026.