Thoughts on AI Phones: Agents Vie for User Entry Point, Memory Becomes Core Asset

At WAIC in July 2026, the Nubia NaviX Ultra, StepFun STEPX Neo, and Honor Robot Phone debuted together, intensifying competition in AI Agent phones. Analysis indicates three routes behind the product forms: vertical integration by traditional manufacturers like Apple, Huawei, and Xiaomi (self-developed hardware, system, accounts, and ecosystem); model companies partnering with ODM to define from scratch (e.g., STEPX Neo, where the model handles intelligence and ODM handles hardware); and deep embedding of model companies with phone brands (e.g., Doubao and Nubia, deeply integrated into system permissions and interaction). These three routes correspond to system-native, from-scratch, and deep embedding, but the core competition always revolves around 'who can become the user's unified entry point.'

The better the Agent, the fewer clicks users make themselves. In the future, the metric for measuring AI phone capabilities may shift from how long users look at the screen each day to how many tasks they delegate to the phone daily. Apps will not disappear, but they will transform from software directly operated by users into capability suppliers behind the Agent. Agents may also become the new generation of search engines and ad slots: when a user requests a restaurant reservation, the system often provides only one result; which platform is called first and which service becomes the default option will affect the final order. Ads may shift from banners and recommendation feeds to the Agent's selection process.

Recently released or announced AI phone solutions all emphasize 'memory.' The Doubao phone assistant can authorize reading recordings, contacts, photos, SMS, notes, and calendar; STEPX Neo establishes separate memory domains for the user and the agent; Huawei's Xiaoyi accesses over 200 system-aware data points and generates memories through 'Xiaoyi Time Machine' by integrating exercise, photos, location, etc.; OPPO launched the 'Xiaobu Next Plan,' opening an on-device Multi-Agent system for beta testing, continuously understanding personal habits. Manufacturers collectively build memory not for a better memo, but because a personal Agent needs at least three types of information: identity (address, preferences, etc.), history (photos, orders, etc.), and rules of conduct (reasons for not taking early flights, etc.). The first two allow the AI to know the user; the third enables the AI to learn to represent the user.

The phone integrates accounts, phone numbers, contacts, location, payments, biometrics, etc., making it the most core digital identity carrier. The Doubao phone assistant combines the AI key with system authentication, distinguishing between 'request' and 'authorization': pressing the key initiates an intent, and the system confirms identity. The complete chain is typically: identity confirmation → preference retrieval → environment sensing → task planning → service execution → key step confirmation → result written back to memory. The phone's role shifts from a tool to the identity carrier, memory container, and action terminal for the personal Agent. But dependency deepens: when switching phone brands, static data can be copied, but the long-trained personal Agent is difficult to fully migrate. The most sensitive asset of AI phones will extend from user data itself to 'another me' shaped by data.

Although manufacturers showcase complex scenarios like travel planning, the fastest-adopted tasks in practice are likely trivial operations like checking packages, hailing rides, and ordering food. The speed of Agent adoption depends on the cost of task failure: a regular app error only affects the current page, but a system-level Agent can read the screen, invoke multiple apps, access personal information, and execute continuously. A single misjudgment could simultaneously impact chat, payments, files, etc. Especially for international flights, large transfers, and medical appointments, errors have more severe consequences.

Security risks cannot be ignored. A July 2026 mobile Agent security study identified two types of attacks: screen perception attacks and channel abuse attacks. Attackers can embed instructions in text or pixels imperceptible to the human eye, inducing the visual model to execute erroneous operations; tests on five mobile Agent frameworks showed that even low-permission malicious apps could hijack Agent behavior. Another study on mobile GUI Agents tested five visual model Agents across 10 apps and 1,111 samples, achieving attack success rates of 23%–30%. On-device execution reduces data leakage but cannot solve screen spoofing, prompt injection, and permission abuse. Whether task records are traceable, key steps require mandatory confirmation, permissions can be granular per task, immediate termination and revocation are supported, and liability attribution are all urgent issues for manufacturers.

When discussing on-device AI, the industry focuses most on NPU compute power, but in the Agent phase, the core issue becomes how to enable the model to continuously understand the screen, retrieve personal data, and invoke system capabilities within limited power, memory, and flash. Apple retains a small model of about 3 billion parameters in the third-generation Apple Foundation Models, while introducing the 20-billion-parameter AFM 3 Core Advanced, but using a sparse architecture that activates only 1 to 4 billion parameters per inference. Model weights are stored in NAND, loading expert modules into DRAM per task; complex reasoning and tool invocation are handled by Private Cloud Compute. This indicates that on-device models will not infinitely stack parameters but follow a 'small model + sparsification + cloud fallback' route. Overall, on-device models handle voice, screen understanding, retrieval, and basic tool invocation, while the cloud handles complex reasoning and long-text generation, evolving into a multi-model collaborative architecture.

At present, AI Agents are still far from being the primary reason for consumers to upgrade their phones; price, camera, battery life, and brand still dominate purchase decisions. For users familiar with smartphones, many Agent features merely shorten operation steps, offering limited convenience. But the same capability holds vastly different value for the elderly, visually impaired, and those unfamiliar with digital products. AI Agents can receive requests in natural language, identify information, invoke services, and complete processes on behalf of users, effectively adding a layer of 'translation' and 'operation' service between complex apps and users. In 2026, Apple announced new accessibility features integrating Apple Intelligence into VoiceOver and Magnifier, providing more detailed environment descriptions and natural language navigation. The most valuable users of AI Agent phones may not be tech enthusiasts but rather those who have been excluded from digital life by app rules.