Google's Gemini 3.8 Audio release turns a familiar voice-agent compromise into a product-level routing decision. Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking belong to the same real-time multimodal family, but optimize different moments: standard Live prioritizes high-volume, latency-sensitive dialogue and can keep a conversation moving while tools run, while Extended Thinking spends more time on multi-step reasoning and can narrate progress before reaching an answer. The important change is not simply a better voice model. It is the appearance of two explicit operating modes inside one agent surface.
That split matters because a useful voice agent cannot maximize immediate turn-taking and deep deliberation on every utterance. A production router can keep greetings, reversible lookups, clarifications, and background tool calls on the fast path, then escalate ambiguous, high-impact, or long-chain tasks to the deliberate path. Short acknowledgements and progress updates become control-plane behaviour: they tell the user that the system changed modes instead of making the call feel frozen.
This extends New Runtime's earlier view of full-duplex voice as a session primitive and conversation-native speech generation. It also turns the ADK voice-agent evaluation loop into a more specific test: evaluate the routing and hand-off policy, not only the model behind each path.
The practical implication is that the evaluation unit becomes the end-to-end policy. Teams should measure time to first useful audio, interruption recovery, tool completion, escalation precision, and whether consequential actions pause for confirmation. A fast model used for the wrong turn can be worse than a slower deliberate model; a deliberate path used too often makes a real-time system feel broken.
The evidence boundary is still narrow: Google's model card and launch article describe the two modes, visual grounding, multilingual dialogue, and concurrent tool use, but do not independently establish production latency, cost, or reliability. The story strengthens if deployed systems publish traces showing that adaptive routing lowers perceived delay without reducing tool completion or interruption recovery. It weakens if mode changes routinely lose context, duplicate tool work, or hide consequential transitions from the user.