Google ADK enables non-blocking voice agents with gemini-3.8-live

Google has published a guide detailing how developers can build responsive voice agents that avoid silence during long-running backend queries.confirmed The open-source Agent Development Kit (ADK) and gemini-3.8-live enable this through non-blocking tools implemented directly in Python.reported

The evidence

According to the official advisory from the Google Developers blog, standard blocking tools cause 13 seconds of dead air if a database query takes that long.reported This silence leads users to assume the call has dropped and hang up. The publication states that switching from a blocking tool to a streaming tool in a benchmark sales-report workflow cut the time to the agent's first words from 14.4 seconds to 1.2 seconds.reported

The decision

The material explains that developers cannot always make backend queries sub-second, so the architectural fix is to keep the agent conversational while tools run in the background. Instead of wiring up custom client-side WebSocket event loops to manage asynchronous tool state, developers can use the ADK to implement non-blocking voice tools natively. The guide highlights two main patterns: streaming tools for operations with meaningful stages and non-blocking FunctionTool definitions for background execution.

The takeaway

Implementing non-blocking tool calls provides zero dead air, allowing the agent to acknowledge requests immediately and keep dialogue alive. Audible progress updates let the agent narrate interim milestones for multi-stage workflows, making waits sound like active progress. Users can also multitask and ask follow-up questions while long-running queries execute in the background. Developers gain granular delivery control, determining when final results enter the conversation via scheduling behaviors such as WHEN_IDLE, INTERRUPT, or SILENT.

Building on this

Developers can construct voice agents that maintain fluid conversations by leveraging the documented capabilities of the Agent Development Kit. By defining functions as AsyncGenerators and using yield instead of return, tools are automatically marked as NON_BLOCKING, providing immediate status acknowledgements to the model. This allows the agent to speak right away with phrases like "Looking into the EMEA figures for you now" while partial updates stream upstream. The material notes that narration may trail the tool by several seconds, but it prevents complete silence. Developers can also configure response_scheduling to handle final results precisely, choosing to interrupt immediately for critical alerts or wait for natural pauses. This architecture supports parallel dialogue where users receive timely answers without stalling the session, addressing the limitation where standard async def functions block the execution loop until a single value is returned.