Google ADK streaming tools slash voice latency from 14.4s to 1.2s with gemini-3.8-live

Google states that its open-source Agent Development Kit, combined with the gemini-3.8-live model, allows developers to build voice agents that avoid conversational dead air.reported In a blog post published on the Google Developers blog, the company outlined how non-blocking tool execution enables agents to remain responsive even when backend queries take significant time to complete.confirmed

The claim

According to the Google Developers blog, standard blocking tools cause voice agents to produce silence equal to the duration of the backend query, such as 13 seconds for a database lookup.reported Google claims that switching to non-blocking patterns, specifically streaming tools or background execution, eliminates this silence. The company reported that in a benchmark sales-report workflow on gemini-3.8-live, changing from a blocking tool to a streaming tool reduced the time to the agent's first words from 14.4 seconds to 1.2 seconds.reported

The response

The source material presents this as a technical guide without external counterpoints or third-party reactions. Google frames the architectural shift as a necessary fix for live voice conversations, where users typically hang up if they encounter extended silence. The company indicates that prompt instructions alone cannot resolve blocking issues because the model's turn is blocked waiting for function response payloads.

What the evidence means

The evidence rests on Google's internal benchmark of a sales-report workflow. The material does not provide independent verification of these timing improvements across different models or use cases. It suggests that the improvement is tied specifically to the implementation of AsyncGenerator functions and the ADK's handling of non-blocking declarations within the Gemini Live API environment.

In practice

Enterprise IT buyers should note that this functionality requires using the open-source Agent Development Kit and the gemini-3.8-live model.reported Developers must rewrite standard async def functions to use yield statements, creating AsyncGenerators that ADK marks as NON_BLOCKING. This allows agents to speak immediately while tools run in the background. Users gain audible progress updates and the ability to ask follow-up questions during long queries. Access is available through the ADK toolkit, but effectiveness depends on implementing streaming or background patterns correctly.