Doubao’s Second AI Phone Bets on Permissioned Agents, Not Screen Control

ByteDance is returning to smartphones through ZTE’s Nubia brand, not with a ByteDance-branded handset. The Nubia NaviX Ultra pairs the Doubao mobile assistant with a new access model designed to avoid the biggest problem facing the December 2025 Nubia M153 preview: an agent that could read screens and imitate taps across third-party apps was powerful, but it collided with app-platform controls, privacy concerns, and authorization disputes. 57

From a sold-out preview to a commercially cleared phone

The Nubia M153 was released as a limited technical preview in December 2025. Its early appeal was straightforward: users could give a voice command and have Doubao attempt tasks such as comparing prices or ordering food across apps. The initial batch of roughly 30,000 units sold out quickly, signaling strong interest in a phone that could act rather than merely answer. 56

The same system-level reach created its central obstacle. Major platforms, including Alibaba, Tencent, and JD.com, restricted the assistant’s access shortly after the preview’s release, undermining the cross-app automation that made the device distinctive. ByteDance subsequently limited the tool in some circumstances, including payments. 5

NaviX Ultra is the attempt to turn that prototype into a viable product. Nubia says the device has received China’s Ministry of Industry and Information Technology network-access license and is due to launch in September. Reports describe that approval as completing the route from large-model filing to terminal network access—an important commercial milestone, though it does not settle whether enough apps will support the agent’s workflows. 37

The key change: app-provided access instead of GUI automation

The M153’s controversial model was essentially GUI automation: with permission, the agent could inspect what was on screen and operate an app through simulated taps. NaviX Ultra is reported to be moving away from that model for major apps. Instead, Doubao will need access to services or interfaces—including MCP-style services—that participating apps choose to expose, with defined data and action scopes. 3

That changes the trade-off.

• For users: a permissioned interface can make it clearer what an agent is allowed to read or do.
• For apps: it gives platforms control over the actions and data they expose to third-party agents.
• For Doubao: it reduces reliance on brittle screen interpretation and tap simulation, but makes the experience dependent on partner integrations.

Nubia has said agent actions will be user-authorized, subject to tiered permissions, and visible and controllable throughout the process. Sensitive actions, including financial transactions and social posting, are intended to require explicit confirmation. 2

Those safeguards matter because a useful phone agent needs access to highly personal territory: messages, calendars, location, documents, accounts, and payments. The practical test is not whether an agent can demonstrate a multi-step flow, but whether it can reliably complete useful flows without obtaining excessive privileges or surprising the user.

What makes NaviX Ultra a “native agent” phone?

Nubia’s claim is not simply that NaviX Ultra has AI features. The company is positioning Doubao as an operating-system-level agent designed to understand a natural-language request and carry out approved tasks across multiple applications. Reported examples include price comparison, ordering, and trip planning. 310

That is a different ambition from familiar smartphone AI features such as transcription, translation, photo editing, search, or a standalone chatbot. A native agent is meant to combine several steps into one delegated task: interpret an objective, select a sequence of actions, request confirmation where necessary, and return with the outcome.

Nubia also says key AI computation and memory are handled on the device. On-device processing can reduce the amount of sensitive material sent to cloud services, but it should be treated as one privacy control rather than a blanket guarantee: the agent’s real privacy posture also depends on permissions, data retention, partner interfaces, and user-facing controls. 12

“World’s first” is a category claim, not a settled fact

Nubia markets NaviX Ultra as the world’s first AI-agent smartphone. That label is contested by the nature of the category itself: other manufacturers and AI companies are also describing devices with system-integrated, task-executing agents as agentic phones. The more meaningful measure will be the available tasks, the number of supported services, and the safety and reliability of completed actions.

NaviX Ultra’s move away from unrestricted GUI control may make it more acceptable to partners and regulators. But it also creates a short-term constraint: an agent can only perform a seamless end-to-end task where the necessary services agree to expose the appropriate capabilities. 35

How StepFun’s StepX Neo compares

StepFun introduced a different version of the same broader idea in July 2026. Its launch combined four pieces: the STEPX hardware brand, Step AOS agent-native operating system, Amoo personal agent, and Step Edge on-device model. 20

StepFun describes Step AOS as a system built from Android, Linux, and RTOS components, with Amoo integrated into the system rather than installed as an ordinary app. Amoo is intended to take a natural-language request and chain work across apps, web tools, and phone functions; StepFun also emphasizes on-device and offline execution for core tasks. 17

The strategic contrast is clear:

Approach	Nubia NaviX Ultra with Doubao	StepX Neo with Amoo
Starting point	An established phone maker working with ByteDance’s Doubao assistant	An AI company building a hardware brand, agent, and agent-native OS together
Current milestone	MIIT network-access approval and a planned September commercial launch	Product and platform unveiled in July 2026
Cross-service model	Reportedly moving toward app-provided, permissioned interfaces	Promises task chaining across apps, web tools, and device functions
Privacy pitch	Local AI processing plus user authorization, tiered access, and confirmation for sensitive actions	Core tasks described as on-device and offline

Both companies are pursuing a phone where the agent becomes the primary way to get work done. Neither can escape the ecosystem question: services that hold payments, identity, communications, and commerce have strong incentives to govern how agents interact with them.

Why phone makers see an opening

Basic generative-AI features are becoming common. Counterpoint forecasts that GenAI-capable smartphones will account for 45% of global shipments in 2026, up from 36% in 2025. As AI becomes standard hardware-table stakes, manufacturers need a more tangible reason for people to upgrade. 37

A capable, trusted agent could be that differentiator. Instead of asking a user to open several apps and repeat the same details, the phone could help assemble a trip, compare options, prepare an order, or coordinate a schedule—while leaving consequential choices to the user. The appeal is not a smarter chat window; it is less manual navigation.

Privacy and user control are equally central to this pitch. Research on agentic AI highlights demand for transparency and fine-grained control around personal and sensitive data. 33 A phone agent becomes persuasive only if a user can understand what it will access, what it will do, when it needs approval, and how to stop or review it.

The bottom line

Doubao’s return through NaviX Ultra is less a repeat of the M153 than a redesign of the agent’s social contract with apps. The first preview proved there was interest in a system-level assistant that could act across services; the second-generation strategy accepts that broad screen control is difficult to sustain in a tightly governed app ecosystem.

Nubia’s commercial and compliance progress is meaningful. Yet the success of NaviX Ultra—and of StepFun’s competing vision—will hinge on a harder product question: can permissioned, auditable integrations deliver enough genuinely useful actions to make an AI agent feel like a new phone experience rather than another AI button? 37