China’s Agentic AI Phones: How AI Agents Are Replacing Apps
TL;DR — At the World Artificial Intelligence Conference in Shanghai in July 2026, three Chinese phone makers unveiled devices that don’t run apps the way every phone has for the last fifteen years — they run AI agents that act across the whole system. ZTE’s Nubia NaviX Ultra has no home screen and no conventional app store; you state a goal by voice or button and its ByteDance-built Doubao agent does it. StepFun’s StepX Neo runs a purpose-built OS that splits the phone into four primitives an agent recombines on the fly. Honor’s Robot Phone adds a mechanical gimbal that nods, shakes its head, and does a backflip to cheer you up. The NaviX Ultra’s first 30,000 units sold out in hours, with resale prices doubling. IDC expects AI-capable devices to make up more than half of China’s smartphone market in 2026. Behind this is a paradigm shift the American incumbents structurally cannot follow: an agent layer that dissolves apps into tasks would demolish the app-store economics that fund Apple and Google. One detail nobody expected — all three phones use Anthropic’s Model Context Protocol to give their agents system-level access. But the euphoric coverage skips the hard part: China’s own super-apps, WeChat and Alipay and Taobao, blocked the agent over fraud and data fears. This guide covers the phones, the strategy, the economics, the behavior change, and the friction the hype leaves out.
China’s Agentic AI Phones Compared
| Phone | Company and agent | Approach | Main limitation |
|---|---|---|---|
| Nubia NaviX Ultra | ZTE with ByteDance’s Doubao | Voice- and button-led phone without a conventional home screen | Cross-app control depends on platform cooperation |
| StepX Neo | StepFun with Amoo | Step AOS recombines system capabilities around each task | Limited ecosystem and interoperability |
| Honor Robot Phone | Honor with YOYO and Alibaba collaboration | Agentic software combined with an expressive mechanical camera | China-focused launch and uncertain global availability |
Key sources: Computerworld on AI replacing smartphone apps in China; Harvard Business Review on China’s AI agents and commerce; and Channel News Asia reporting from WAIC 2026.
Related analysis
- McKinsey AI 2026 Trust Report on AI Agents
- Claude Cowork and Autonomous AI Agents
- The Rise of Agentic Engineering
- AI News Roundup: Agentic Systems and Model Releases
- ByteDance, TikTok and AI Platform Strategy
The Shift: From Apps You Operate to Agents That Act
The defining idea is architectural, and it’s easy to miss under the marketing. These aren’t phones with an AI app added. They’re phones where the operating system itself has an agentic layer that works across apps and, increasingly, instead of apps.
For fifteen years the smartphone interaction model has been constant: you decide which app performs a task, open it, navigate its menus, complete the task, then switch to the next app and repeat. AI, in this model, is just one more app — you open ChatGPT or Doubao the way you open any other icon.
The Chinese agentic phones invert this. You state a goal — “book me a table near the office at 7 and message the group” — and the agent interprets the intent, works across whatever apps and services it needs, and completes the task, often with no screen interaction at all. The agent isn’t an app on the phone. The agent is the interface.
Nubia’s chief Ni Fei drew the distinction sharply at the launch: many so-called AI phones simply stack AI functions on top of an existing system, which he argued actually makes them more cumbersome to use, not less. The point of the redesign is to remove the app-by-app navigation entirely.
StepFun’s architecture is the clearest illustration. Its Step AOS — built on Android, Linux, and a real-time OS — deliberately breaks the phone down into four primitives: communication, apps, files, and system tools. The built-in agent, Step Amoo, recombines those primitives dynamically based on the user’s stated goal. Instead of “open the messaging app, then the calendar app, then the maps app,” the agent treats all of the phone’s capabilities as a single toolbox and assembles the right combination itself. It’s a genuinely different mental model of what a phone is.
Three Phones, Three Strategies
Each of the three flagship devices takes a distinct approach — mass market, architectural purism, and emotional attachment — which together map the range of what an agentic phone can be.
Nubia NaviX Ultra (ZTE) — the mass-market first mover. Running ByteDance’s Doubao, China’s most popular AI assistant, the NaviX Ultra is the one that proved demand. It has no home screen and no conventional app store; interaction happens by voice or a dedicated button. Prototyped in December at 3,499 yuan (about $516), its first batch of 30,000 units sold out within hours, and resale prices doubled on secondary markets. ZTE billed it as the world’s first mass-produced agentic AI phone. (Worth noting: ZTE is banned from the US by the FCC on national-security grounds, so this device will never officially reach American consumers.)
StepX Neo (StepFun) — the architectural purist. StepFun was founded in 2023 by Jiang Daxin, a former Microsoft vice president and chief scientist at Microsoft’s Software Technology Center Asia. Its StepX Neo runs the proprietary Step AOS with the Amoo agent built into the OS. The device includes a “Soul Configuration” for defining the agent’s persona, a user-identity layer so the agent knows who it’s serving, and explicit AI-agent security controls for managing permissions. This is the most ground-up reimagining of the three — an agent-native OS rather than an agent layer over Android.
Honor Robot Phone — the “Attachment Economy” play. Honor, spun off from Huawei, took the most unexpected direction. Its Robot Phone, built on MagicOS 10 (which Honor now calls “Agentic OS”), pairs an agent co-developed with Alibaba (Yoyo) with a mechanical gimbal holding a 200-megapixel camera. The camera can track a moving subject while the phone sits on a table, or stabilize footage while you walk. But the real novelty is expressive: the phone nods to say yes, shakes its “head” to say no, and does a backflip to cheer you up. Honor plans to launch it August 12 in China. It’s the clearest example of what some call the “Attachment Economy” — human-like traits designed to make people emotionally bond with a device that seems to have thoughts and feelings.
The American company nobody expected in this story. All three phones rely on Anthropic’s Model Context Protocol (MCP) — the same open standard behind much of the Claude ecosystem — to give their agents system-level access to apps and data. In a story about Chinese independence from American mobile platforms, the connective tissue underneath the agents is an American open standard. It’s a reminder that the infrastructure layer of AI is more entangled across the US-China divide than the geopolitics suggests.
Why China Can Blow Up the App Model — and Apple and Google Can’t
The most important strategic fact about agentic phones is that the incumbents are structurally prevented from following. The economics that make Apple and Google powerful are the economics an agent layer destroys.
Consider what apps are worth to the platform owners. Apple takes 15-30% of app revenue (cut to 25% in mainland China as of March 2026). The Play Store, by estimates introduced during the Epic v. Google litigation and not confirmed by Google, has historically accounted for somewhere between 17% and 26% of Google’s operating income. The entire economic logic of iOS and Android rests on apps as the unit of commerce — discrete, downloadable, monetizable, and taxable at the store.
An agent layer that dissolves apps into tasks doesn’t just threaten that model; it dismantles it. If the agent books your restaurant, orders your groceries, and manages your subscriptions without you ever opening an app or passing through a store, the 15-30% cut has nowhere to attach. Apple and Google will not voluntarily let a third-party agent replace the app layer in their operating systems, because doing so would demolish the revenue that funds everything else they build. They may add agentic features that cross app boundaries — cautiously — but handing the interface to an agent that makes their app store irrelevant is not a move they can make.
This is the asymmetry that defines the race. The Chinese OEMs had no app-store profits to protect. ZTE, StepFun, and Honor never built the iOS/Android app economy and captured none of its rents, so they lose nothing by burning it down. They can pursue the agentic model precisely because they have no incentive not to — a textbook case of the innovator’s dilemma, where the incumbents’ greatest strength becomes the exact thing that immobilizes them.
The timing is driven by desperation as much as vision. China’s smartphone shipments have fallen for five consecutive quarters, squeezed by a memory-component cost crisis and weak consumer demand, and IDC expects the global market’s steepest annual decline on record in 2026. In a commoditized, shrinking market, agentic AI is the escape hatch — a reason to buy a new phone and a way to differentiate from Apple without competing on the same terms. Necessity is doing a lot of the work here.
The Behavior Change: What Happens When the Agent Does the Choosing
The shift isn’t only in how phones work — it’s already changing how hundreds of millions of people buy, browse, and interact, and the early data shows the app economy eroding from the inside.
The clearest signal is financial: in-app purchases fell roughly 40% by early 2026, even as the average spend per purchase rose (because what remains is dominated by subscriptions to AI and “vibe-coded” apps). The old model of downloading an app, browsing it, and buying inside it is contracting. Even inside the app stores, the transaction is shifting from apps to AI replacements.
The scale of the behavioral shift is striking. More than 600 million people in China are estimated to have used some form of agentic app. This is where China’s structural advantage becomes decisive: the country entered the AI era with super-apps — WeChat, Alipay, Taobao, Meituan, Douyin — already unifying messaging, payments, commerce, and services into single platforms embedded in daily life. When Alibaba or ByteDance inserts an agent into that flow, it doesn’t need to convince anyone to adopt a new tool; it upgrades one they already use every day.
Meituan’s Xiaomei agent, launched in late 2025, illustrates the delegation model. Executives described it internally not as a chatbot but as an “orchestrator plus execution agent.” A user says, “Order my usual lunch, but deliver it 20 minutes later today,” and the agent interprets the intent, applies stored preferences, and completes the transaction — often with zero screen interaction. The point isn’t convenience; it’s delegation. The user stops operating the phone and starts directing it.
This compounds over time in a way that matters competitively. The platform with the most complete transaction data trains the most accurate agent. China’s super-apps entered 2026 already holding unified commerce, payment, and messaging data; Western AI platforms are still negotiating access to the transaction layer app by app. The behavioral shift and the data advantage reinforce each other.
What it means for the phone as an object: the screen is becoming less central. When the agent does the choosing and the transacting, the grid of app icons — the organizing metaphor of the smartphone since 2007 — starts to feel vestigial. Honor’s Robot Phone points even further out, toward a device with a body and personality rather than just a glass rectangle. Whether people want that is an open question, but it’s a genuine reimagining of what the object is for.
What It Means Globally — and the Friction the Hype Skips
The agentic phone is a real paradigm shift with real geopolitical weight. It’s also less finished than the launch-day sellouts suggest, and honesty about both halves is what separates analysis from cheerleading.
Why it matters beyond China:
Independence from Android. Google’s services layer already vanished from China years ago. Now the Android app layer — the last structural piece of Google’s ecosystem operating in the country — is being replaced by vertically integrated agent stacks owned by Chinese super-app companies. This is the completion of a long decoupling.
An export of worldview. If agentic phones succeed domestically, they’ll be offered internationally — and selling the hardware means exporting Chinese super-apps, Chinese financial services, and AI-shaped information. An agent reflects the values, assumptions, and constraints of whoever built it. Adopting a Chinese agentic phone means, to some degree, adopting the worldview baked into its agent.
The Attachment Economy. Honor’s gesturing, backflipping phone is an early instance of a broader pattern: using human-like traits to make consumers emotionally attached to devices that appear to have inner lives. The coming wave of AI wearables — glasses and watches that converse with your phone, which converses with AI models — will deepen this. The question of whose values those models reflect becomes more pressing as the devices become more intimate.
A fork in the road. The blunt framing is: the world will choose between the Chinese agentic-AI model and the American app model. That may be too binary — hybrid outcomes are likely — but the divergence is real, and it’s the clearest fork in mobile computing since iOS and Android split the market.
The friction the euphoric coverage skips:
The super-apps fought back. This is the detail that reframes everything. When the Doubao phone’s agent — which reads screens and acts like a user — tried to operate across apps, major platforms including Taobao, Alipay, and WeChat blocked it, citing fraud and data-exposure risks. An agent that acts across apps needs those apps to cooperate, and China’s most powerful app owners have every incentive not to hand a rival’s agent access to their walled gardens. The agentic vision runs directly into the super-app owners’ control of their own ecosystems.
MCP-only access is limited. Because GUI-based control triggers exactly the resistance above, mass-produced phones lean on MCP: the agent can only reach apps that expose an MCP service and choose to open their data. Most of the ecosystem hasn’t done this yet. The “agent does everything” promise is bounded by how many app owners agree to be reachable.
Fragmentation is chronic. Every OEM has built a proprietary agent layer. Because China blocks Google’s unifying services, switching phone brands can mean switching assistant, app store, cloud, and push notifications all at once. Developers must adapt software for each manufacturer’s proprietary layer. Interoperability — the thing agentic AI most needs — is weak, and a standards battle is unfolding to fix it.
The surveillance question. A system-level agent, by design, sees everything you do on your phone. In a market with mandatory data-sharing obligations for companies, an agent with that level of visibility is a live governance concern, not a hypothetical one — and one reason Western regulators will scrutinize these devices closely if they arrive.
The balanced take: the agentic phone is genuinely a new paradigm, and China is shipping it first because it has the super-app data, the OEM freedom from app-store economics, and the market pressure to justify the gamble. All of that is real. But “the agent handles everything” remains more demo than daily reality, because the app owners hold veto power over agent access and are actively using it. The metric to watch isn’t launch-day sellouts — those measure novelty demand. It’s adoption depth: how many tasks people actually delegate, how many apps actually open their data, and whether the agent becomes the interface or stays a party trick.
Frequently Asked Questions
What is an agentic AI phone?
An agentic AI phone is a smartphone whose operating system has a built-in AI agent that can act across apps — and increasingly instead of apps — to complete tasks autonomously. Rather than opening individual apps and navigating their menus, the user states a goal by voice or button, and the agent interprets the intent and executes it, often with no screen interaction. This differs from ordinary phones that simply add an AI assistant as one more app.
Which Chinese companies are making agentic AI phones?
Three led the July 2026 World AI Conference launches: ZTE (Nubia NaviX Ultra, running ByteDance’s Doubao agent), StepFun (StepX Neo, running its proprietary Step AOS with the Amoo agent), and Honor (Robot Phone, running MagicOS 10 with a Yoyo agent co-developed with Alibaba). All three use Anthropic’s Model Context Protocol to give their agents system-level access.
What is the Nubia NaviX Ultra?
It’s ZTE’s agentic phone, billed as the world’s first mass-produced AI agent smartphone. It has no home screen and no conventional app store; interaction is by voice or a dedicated button, powered by ByteDance’s Doubao agent. Prototyped in December 2025 at 3,499 yuan (about $516), its first 30,000 units sold out within hours, and resale prices doubled. ZTE is banned in the US by the FCC, so it won’t officially reach American consumers.
What makes the Honor Robot Phone different?
The Honor Robot Phone adds a mechanical gimbal holding a 200-megapixel camera that can track moving subjects and stabilize footage. Its distinguishing feature is physical expressiveness: it nods to say yes, shakes its “head” to say no, and does a backflip to cheer the user up — an example of the “Attachment Economy,” where human-like traits make people emotionally bond with devices. It launches August 12, 2026 in China.
Why can’t Apple and Google make agentic phones like this?
Because the economics prevent it. Apple takes 15-30% of app revenue and the Play Store historically accounted for an estimated 17-26% of Google’s operating income. An agent layer that dissolves apps into tasks removes the app-store transaction that both companies tax, dismantling their revenue model. The Chinese OEMs never captured app-store profits, so they lose nothing by replacing the app model — a classic innovator’s dilemma.
How is AI changing app usage in China?
In-app purchases dropped roughly 40% by early 2026 as agents began replacing app-by-app interaction, though average spend per purchase rose because AI and subscription apps now dominate. More than 600 million people in China have used some form of agentic app. Super-apps like WeChat, Meituan, and Taobao are embedding agents that complete tasks — ordering food, booking travel — with little or no screen interaction, shifting users from operating their phones to directing them.
What is Anthropic’s Model Context Protocol and why does it matter here?
The Model Context Protocol (MCP) is an open standard, developed by the American company Anthropic, that lets AI agents access external apps, tools, and data in a structured way. All three Chinese agentic phones use it to give their agents system-level access. It’s notable because a story about Chinese independence from US mobile platforms relies on American-originated infrastructure at the agent layer.
Did the agentic phones actually work as advertised?
Partially, and with significant friction. When the Doubao phone’s agent tried to operate across apps by reading screens and acting like a user, major platforms including Taobao, Alipay, and WeChat blocked it over fraud and data-exposure concerns. Mass-produced phones therefore rely on MCP, which only reaches apps that voluntarily expose their data — and most haven’t yet. The “agent does everything” promise is real in demos but bounded in daily use by app owners’ cooperation.
Will these phones come to the US or Europe?
Not in their current form. ZTE is banned in the US by the FCC on national-security grounds, and the StepFun and Honor devices are built for the Chinese market with no US versions announced. If the model succeeds domestically, international versions are likely — but they would face regulatory scrutiny in Western markets, particularly over the surveillance implications of a system-level agent and the export of Chinese super-app ecosystems.
What does this mean for the future of smartphones globally?
It represents a genuine fork in mobile computing: a Chinese agentic-AI model where the agent is the interface, versus the American app model where apps remain the unit of interaction. The divergence will likely deepen with AI wearables — glasses and watches that route intelligence through the phone. The open question is whether the world prefers delegating to an agent or operating apps directly, and whose values the agents doing the delegating reflect.
Final Take
The most consequential thing about China’s agentic phones isn’t the hardware — it’s the demonstration that the app, the organizing unit of mobile computing since 2007, is not permanent. For fifteen years “using your phone” has meant choosing and operating apps. The Chinese OEMs are betting it will come to mean stating a goal and letting an agent handle the rest, and they’re shipping that bet at scale while Apple and Google watch from inside an economic model that won’t let them respond.
That structural asymmetry is the real story. This isn’t China being more innovative in the abstract; it’s China being free to be innovative here because it holds none of the app-store rents that immobilize the incumbents. Combined with super-apps that already unified the transaction layer and a domestic market desperate for a reason to upgrade, the conditions for this specific leap existed in China and nowhere else.
But the honest version includes the friction. The super-app owners blocked the agent. MCP access reaches only the apps that agree to be reached. Fragmentation makes the ecosystem incoherent. And a system-level agent that sees everything raises surveillance questions that will follow these devices anywhere they try to go. The launch-day sellouts measure curiosity, not transformation. What will actually tell us whether the smartphone is being reinvented is adoption depth — how many tasks people truly delegate, and how many app owners open their doors to someone else’s agent.
For everyone outside China, this is a preview worth studying closely rather than dismissing or over-hyping. The agentic interface is coming to phones, wearables, and everything downstream of them. The question the rest of the world will eventually have to answer is the one this launch poses: do you want to operate your technology, or direct it — and if you direct it, whose agent are you trusting to act on your behalf?
Published August 2026 · The AI & Tech Society · digitalstrategy-ai.com
Sources: Computerworld (“How AI in China is killing smartphone apps”); The Next Web, Sixth Tone, Business Upturn, and BigGo Finance coverage of the WAIC 2026 launches; Cobo and Macrostream reporting on the NaviX Ultra, StepX Neo, and Honor Robot Phone; IDC smartphone market forecasts; The Economist and Straits Times (“AI super-apps are remaking China’s internet”); PYMNTS and CNBC on super-app agentic commerce; Harvard Business Review (“What China’s AI Agents Reveal About the Future of Commerce“) on Meituan’s Xiaomei agent; Lawfare/Hans India reporting on super-apps blocking the Doubao agent; Apple China commission-rate reporting. App-store revenue percentages for Google reflect estimates introduced during Epic v. Google litigation and not confirmed by Google. Product launch dates, prices, and specifications are as reported at WAIC 2026 and subject to change. Verified August 2026.
Discover more from The Tech Society
Subscribe to get the latest posts sent to your email.