The thesis is: responsive voice agents need tool execution that can continue without blocking the conversational loop. Rather than treating silence as a prompting problem, the guide locates the issue in how tool work is scheduled and reported.
The mechanism is: ADK streaming tools yield progress updates while work continues, whereas background execution suits tasks where repeated narration would be intrusive. Response scheduling then controls how the eventual result re-enters the conversation. Google reports that in its sales-report workflow on Gemini 3.8 Live, replacing a blocking tool with a streaming tool reduced time to first words from 14.4 seconds to 1.2 seconds. The why now is: live voice exposes backend latency as conversational dead air, making execution design directly perceptible to users.
The first implication is: Teams building voice agents can test asynchronous tool patterns against their own slow queries instead of relying on prompts to mask waits. The evidence boundary is: the latency comparison is a Google-reported benchmark for one workflow, not an independent or general performance result. What to watch: whether ADK implementations handle duplicate tool calls, cancellation, and delivery timing reliably under real workloads. The guide is implementation guidance, not evidence that these patterns improve every voice application.