Welcome back to LLM Decode 👋
AI is quickly moving beyond the familiar type a prompt, get an answer experience. Google’s newest Gemini models are bringing deeper reasoning into live voice conversations, while Huawei predicts autonomous agents could eventually generate the overwhelming majority of AI activity. Together, they point toward an internet where we increasingly talk to AI, and AI increasingly acts for us.
Here’s what matters today.
Gemini Can Now Reason While You’re Talking
Google has introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its newest models designed specifically for real-time voice interactions. Google says they bring near-real-time reasoning to voice agents, allowing conversations to feel more natural while AI handles increasingly complex tasks.
The two models target different workloads. Gemini 3.8 Live prioritizes low latency, scale and cost efficiency, while Extended Thinking is designed for harder requests that require deeper, multi-step reasoning in the background. Both support native audio interactions, while Gemini 3.8 Live also supports asynchronous function calling so tools can run without necessarily stopping the conversation.
Google is also pushing the technology beyond developers. Gemini 3.8 Live is powering Search Live, while the broader technology is being integrated across experiences including Gemini and Google Workspace.
Why it matters
Voice is becoming an interface, not just an input. AI can increasingly reason, use tools and continue working while maintaining a conversation.
Businesses can build voice-first agents. Customer support, sales qualification, scheduling and internal assistance are obvious early applications.
Different reasoning levels could reduce costs. Routine conversations can use faster models while difficult tasks trigger deeper reasoning.
Typing may become optional for more workflows. The bigger shift is from chatting with AI toward working with AI through conversation.
Huawei Thinks AI Agents Could Generate 90%+ of AI Traffic
Huawei has made a striking prediction: by 2035, autonomous agents could account for more than 90% of global AI token traffic. The forecast appears in its new Intelligent World 2035 report.
Why so much activity? Unlike a chatbot waiting for a human prompt, agents can continuously perceive information, reason, make decisions, call tools and interact with other systems. Huawei projects that annual global token consumption could grow 100,000-fold by 2035 as these systems proliferate.
But Huawei's numbers are forecasts, not established outcomes. The company argues that an agent-driven world would require dramatically more computing and networking infrastructure, alongside stronger systems for keeping autonomous software secure, private and controllable.
Why it matters
AI usage could become machine-driven. Today humans initiate most AI interactions. Agents could eventually generate enormous amounts of activity themselves.
Token economics become critical. Businesses deploying thousands of agent actions will need to monitor cost per completed task, not simply API price.
Security changes with autonomy. Giving agents access to browsers, databases, payments or business software creates a much larger permission problem.
Infrastructure becomes part of the agent race. More autonomous work means substantially more demand for compute, networking and cloud capacity.
4 AI Tools & Community Workflows
Voice-to-Action Workflow: Instead of using voice AI only for Q&A, test Speak → Understand → Call Tool → Complete Task → Confirm. Start with meeting scheduling, CRM updates or internal information retrieval.
Reasoning Router: Don't use deep reasoning for every request. Route simple conversations → fast model and complex decisions → extended-thinking model to balance speed, quality and cost.
Agent Cost Dashboard: Track tokens consumed → tools called → time saved → successful tasks. As agents become more autonomous, measuring cost per completed outcome becomes more useful than simply measuring prompts.
Agent Permission Map: Separate tools into Read / Recommend / Act with Approval / Act Automatically. Give new agents minimal permissions first and expand them only after testing.
Ready to level up your AI skills?
Explore Our Courses →That’s it for today.
The AI space doesn’t slow down - and neither should your thinking.
See you in the next drop.