---
title: "Gemini 3.8 Live does not end when the model stops speaking. Voice agents need a second state machine."
description: "Google's extended-thinking Live model can finish an utterance while reasoning or tool calls continue in the background. Existing voice clients that equate turn completion with idle state can now interrupt work that is still running."
url: "https://devencelab.com/insights/2026/09/16/gemini-3-8-live-does-not-end-when-the"
date: "2026-09-16"
section: "Insights"
tag: "Agentic AI"
author: "Devence Lab"
reading_time: "2 min read"
site: "Devence Lab"
license: "Readable and quotable with attribution to the canonical URL."
---

# Gemini 3.8 Live does not end when the model stops speaking. Voice agents need a second state machine.

Google's extended-thinking Live model can finish an utterance while reasoning or tool calls continue in the background. Existing voice clients that equate turn completion with idle state can now interrupt work that is still running.

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September. The important production change is not better speech. In the extended-thinking model, `turnComplete: true` can arrive while background reasoning or tool execution is still in progress.

That breaks a common voice-agent assumption: when the model stops speaking, the turn is over. Google's API now separates utterance completion from interaction completion, which means client state has to do the same.

## Silence and idle are now different states

Google documents a new `interaction_status` signal for extended-thinking sessions. `IN_PROGRESS` means reasoning, processing or asynchronous tool work may still continue; only `IDLE` means the interaction is finished and the system is ready for a new turn.

The difference matters because the model can speak an intermediate update, mark that utterance complete, call a tool in the background, then speak again. A client that treats the first `turnComplete` as idle can reopen input, change interface state or submit new content while the original task is still active.

> A voice agent now needs to track whether the model finished speaking and whether the task finished executing as separate facts.

This is not cosmetic session bookkeeping. Google states that sending client content with `turn_complete=true` interrupts active generation. A stale state transition can therefore become a control-flow error rather than merely a confusing interface moment.

## Extended thinking also changes the tool contract

Gemini 3.8 Live Extended Thinking supports asynchronous function calling only. Google requires tools to use non-blocking behaviour; synchronous blocking declarations return an error. Teams migrating an existing Live integration therefore need to review both the client lifecycle and every tool declaration.

That creates two independent migration checks. First, replace idle decisions based solely on `turnComplete` with `interaction_status`. Second, confirm that tool orchestration can tolerate responses arriving after intermediate speech rather than assuming one request, one tool phase and one final utterance.

## Test interaction completion, not just response latency

Production tests should record utterance completion, interaction status, tool start, tool completion and interruption events on the same timeline. Scenarios need to include slow tools and user speech arriving while background work remains active.

The deployer change is straightforward: treat spoken output as one stream and task execution as another. Gemini 3.8 Live Extended Thinking makes that separation explicit in the protocol. Voice-agent clients that keep a single turn-state flag are now carrying an architectural compatibility risk.

## Sources

- [Thinking in the Live API](https://ai.google.dev/gemini-api/docs/live-api/thinking) - Google AI for Developers
- [Gemini 3.8 Live Extended Thinking](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-live-extended-thinking) - Google AI for Developers