When Your Phone Starts Doing Things for You: The Dawn of AI Agent Smartphones at WAIC 2026

If there was one thing at WAIC 2026 that genuinely excited everyday consumers, it was the arrival of AI agent smartphones.

On July 19, ZTE Nubia gave the NaviX Ultra its first public hands-on debut at the WAIC exhibition floor. Almost simultaneously, Chinese AI unicorn StepFun unveiled its own autonomous smartphone, the STEPX Neo. Two different products, one unmistakable direction: phones that stop being tools you operate, and start becoming assistants that get things done for you.

Two Products, One Signal

NaviX Ultra positions itself clearly: the world's first mass-produced AI agent smartphone. It runs Doubao Phone Assistant, a registered on-device large language model, with all AI inference happening locally. A dedicated physical AI button on the side is its signature design feature — one press wakes the agent, ready for your command.

In a live demo, a user simply said, "Book me a flight to Beijing for tomorrow morning, window seat, pay with Alipay." The phone autonomously opened the travel app, searched for flights, filtered for window seats, and called Alipay to complete payment — without a single manual tap.

STEPX Neo takes a slightly different route. StepFun built it around the Step Agentic-native OS, an operating system that is not app-centric by design, but intent-centric. Chairman Yin Qi made a telling remark at the launch: "This device officially closes the loop — from foundation models and agent systems all the way to the hardware terminal."

What Changed Under the Hood

Over the past two years, phone makers have tried many "AI phone" concepts. Most involved cramming a chatbot into the system or adding a few AI features to the camera and photo gallery. But these were essentially patches on a traditional smartphone.

What this new generation of AI agent phones does is fundamentally different. The tech stack has three layers:

System-level GUI Agent architecture. Instead of simulating finger taps, these agents use MCP (Model Context Protocol) and A2A (Agent-to-Agent) protocols to communicate directly with third-party apps at the service layer. Booking a hotel no longer means opening the app and filling out forms — the agent calls the hotel's service interface directly, yielding an order-of-magnitude efficiency gain.

On-device large language models. NaviX Ultra's Doubao Assistant runs entirely locally. Photos, conversations, usage patterns — none of it leaves the phone. This addresses the single biggest trust barrier for AI phones.

Long-term memory and preference learning. The agent remembers your flight preferences, hotel brands, and payment methods. Repeat requests don't require repeat instructions.

Privacy Is Not Optional

Both products have made explicit privacy commitments. NaviX Ultra emphasizes that "all model inference runs locally on-device, with no data uploaded to the cloud," backed by full generative AI regulatory registration. StepFun has also published data protocol frameworks with its first ecosystem partners, including Trip.com and Alipay.

This isn't marketing fluff. The maturation of on-device inference means that for the first time, privacy and intelligence can coexist without compromise.

A Paradigm Shift: From App-Centric to Intent-Centric

For 17 years, the smartphone interaction paradigm has barely changed: unlock, find an app, tap in, operate, exit, find the next app. AI agent phones are trying to break this loop.

The user expresses an intent — "I'm going on a business trip to Beijing" — rather than an action sequence — open the travel app, search Beijing, filter by time, select a seat, pay. The agent handles decomposing the intent into actions; the user only confirms key decision points.

What does this mean? For everyday users, the barrier to using a phone drops dramatically. For developers, the logic of app competition changes — what determines user choice is no longer icon placement or home screen exposure, but whose agent capabilities are stronger and whose APIs are more AI-friendly.

Year One

The simultaneous debut of two mass-produced AI agent smartphones at WAIC 2026 sends a clear signal: the leap from AI that "talks" to AI that "does" is not just happening in the cloud and in enterprise scenarios — it is entering the device in everyone's pocket.

Of course, this is only the beginning. First-generation products will inevitably have rough edges, and the depth and breadth of ecosystem partnerships are far from ideal. But the direction is right. When your phone truly understands your intent and gets things done for you autonomously, the interaction logic we've been using for over a decade may finally be ready for retirement.