Google added native live evaluation to the Agent Development Kit on 24 August. A simulated user drives a live voice agent by speaking its turns as audio, the spoken replies are scored, and the whole thing runs inside the same evaluation loop teams already use for text agents rather than in a separate voice harness.
The failure modes named in the post are what makes this more than a convenience. A live agent has to take the right actions across a spoken conversation turn after turn, where timing and recovery matter as much as content, and behaviour that sounded perfect yesterday can change quietly on the next prompt tweak or model iteration: tools stop firing, context slips between turns, interjections go ignored. None of those are audible in a single-turn demo, which is why a demo is not evidence.
Turning the conversation into a repeatable case is the same production move as LangChain turning agent evals into a spec-to-task pipeline, and it is the direction described in evals moving from leaderboards to operating diagnostics. The agents hub collects both.
This is a how-to post with a worked example: it documents the loop without publishing scoring rubrics or reliability data for the simulated user itself. The condition to watch is whether simulated-user audio becomes reproducible across runs, because an eval whose synthetic caller varies between executions moves the noise rather than removing it.