Build voice agents that can respond quickly or think longer without changing your stack. Gemini 3.8 Live is available today in LiveKit Agents. Choose gemini-3.8-live for low-latency audio or gemini-3.8-live-extended-thinking for async reasoning. Check the docs for more info:

Awesome way of visualising the difference! @codeSTACKr's demos are always fire 🔥 Thanks so much for shipping on Day0 🚀

🔥🔥

Fast and deep reasoning in one stack is the right split. latency budgets finally get a product knob.

Having both low latency and extended thinking options without changing the stack is a great flexibility win

Great flexibility fast or deep thinking, same stack.

This is awesome! Thanks for the demo :)

Voice agents should separate speaking state from action state during tool calls.

Keeping the stack fixed should make this a useful regression comparison. I would separate the silence timeout from the tool-execution deadline, then test both models with a slow tool and a caller correction. A quiet audio stream must not trigger a duplicate task or premature success. My detailed QA approach:

extended thinking for voice agents??

the interesting question is whether you can switch modes mid-call or it's locked in per session. real conversations don't sort cleanly into fast vs slow upfront, the hard part is deciding on the fly whether this specific question is worth eating the extended-thinking latency for.

On a phone line turnComplete is not idle. Extended thinking keeps speaking fillers while interaction_status stays in progress, so the next generation still needs to know the turn is not over.