This marks my 425th column article.

On July 14, NetEase Cloud Music's emotional AI offering, 'Miaoshi,' officially discontinued its services.

Tomorrow, July 15, marks the official implementation of the 'Interim Measures for the Administration of AI Personified Interactive Services' (hereinafter referred to as the 'Measures'), coinciding with the offline adjustment of agent functionalities in Doubao and Qianwen.

Shortly after, on July 17, 2026, the World Artificial Intelligence Conference will commence in Shanghai. Ni Fei, Senior Vice President of ZTE Corporation, has announced that Nubia will unveil the world's first AI agent phone, a mass-produced flagship model, at this event.

In just one week, agents bid farewell to app platforms but make a grand entrance into smart hardware. This transition is not merely a change in medium but a profound integration with a 'physical cheat code': evolving from pure digital AI agents to AIoT agents endowed with physical perception and execution capabilities.

On one hand, leading software platforms are distancing themselves from agent ecosystems; on the other, terminal hardware manufacturers are embracing them as a core feature. This divergence is not coincidental but a strategic choice driven by business imperatives and compliance demands.

Behind this contrast lie two pivotal questions for the industry: Why must agents exit general-purpose apps? And, once integrated into hardware as AIoT agents, do they evade the regulatory constraints of personified services?

This article aims to address these questions sequentially.

Survival Imperative: Why Are Apps Phasing Out 'Open Personas'?

First, let's dispel a common misconception: This round of shutdowns does not signal a complete retreat of agents.

A closer look at the July 4 announcements from Doubao and Qianwen reveals a specific target: 'user-created agents,' which empower users to customize the personality, rules, and interaction logic of AI entities. Platforms are merely discontinuing the open production of personas.

Why? Three critical factors are at play.

The first is the computational cost.

Volcano Engine reported in late June that Doubao's large model exceeded 180 trillion daily Token usage, a 1,500-fold surge from May 2024. By month-end, Doubao introduced paid subscriptions, ranging from 68 to 500 yuan. Despite this, 'LatePost' reports indicate that Doubao's daily revenue in the first half of 2026 was less than 1 million yuan, while computational costs soared to tens of millions. Estimates suggest that an agent's Token consumption for complex tasks can be 5 to 10 times that of a standard conversation. Unlike traditional internet platforms where increased users reduce marginal costs, large model products face escalating costs with user engagement.

The second is content moderation.

Open-produced personas are boundless. Internet analyst Zhang Shule likened user-generated content to 'wildfire.' During Shanghai's 'Clean Internet' campaign, over 14,000 non-compliant agents were removed. Post-implementation of the 'Measures,' each open-produced persona will necessitate content review, algorithm filing, and minor protection measures. Compliance obligations grow linearly with the number of personas, without a corresponding revenue increase.

The third is permission limitations.

Agents operating within app frameworks lack system-level permissions and must rely on 'screen reading + simulated clicks' to interact with other apps. This approach is both fragile and risky, susceptible to system blocks or UI update invalidations.

Collectively, these factors lead to a clear conclusion: Apps, as traffic containers, cannot sustain agents.

Agents are a new breed that incurs costs with each operation and accumulates compliance obligations with each persona. They urgently require a new habitat.

Compliance-Driven Migration: Emotional Agents to Licensed Apps, Task Agents to Hardware

This migration follows two distinct paths, rooted in Article 2 of the 'Measures.'

The regulation targets 'continuous emotional interaction services that simulate human personality traits, thinking patterns, and communication styles' but explicitly exempts 'services providing intelligent customer support, knowledge Q&A, work assistants, learning and education, scientific research, etc., that do not involve continuous emotional interaction.'

Thus, emotional agents migrate to licensed vertical apps, such as ByteDance's CatBox, which implements stringent review and anti-addiction systems. Task-oriented agents, meanwhile, transition to hardware, marking the onset of AIoT agent evolution.

This directly answers the initial question: Why does the agent phone dare to debut prominently? Because its core functionalities—ticket booking, price comparison, ordering, and cross-app operations—fall within the exemption list. It avoids personification regulatory requirements by virtue of its service type.

More importantly, hardware provides task agents with five advantages that apps cannot offer:

1. Monetization. By late 2025, ByteDance's Doubao and ZTE launched the Nubia M153 engineering prototype, priced at 3,499 yuan. The initial batch of about 30,000 units sold out immediately, with some commanding premiums on secondhand platforms. Users may hesitate to pay monthly fees for virtual conversations but are willing to pay upfront for hardware that 'works for them.' Headleo Research's 2025 AI toy industry report confirms this trend: LOVOT charges 880 yuan monthly, and BubblePal has a 40% repurchase rate. The 'hardware + subscription' model has proven successful in other categories.

2. Permissions. System-level permission delegation is a one-time process. The M153 employs tiered permission management, requiring mandatory manual confirmation for sensitive operations. Authorization shifts from cumbersome negotiation in each conversation to a one-time contract at purchase.

3. Identity. Regulation requires a tangible entity. Millions of user-created personas are untraceable, but a hardware product has a name, a filing entity, and a recall mechanism.

4. Memory. On-device storage retains long-term memories locally, reducing privacy risks associated with constant cloud data transmission.

5. Physical interaction (the essence of AIoT). Pure software agents are confined to screens, while AIoT agents perceive the world through sensors and interact with the physical environment through hardware. Hardware becomes an extension of the agent's senses and limbs into the real world.

While traffic containers cannot balance the three critical factors, 'billing containers' can manage two and a half. The remaining half reveals a fundamental conflict between AI hardware and internet giants.

By late 2025, leading apps like WeChat, Taobao, and Meituan imposed restrictions on Doubao AI phones. Hardware controls the system but not the ecosystem. From the giants' perspective, Doubao's GUI-based agent approach is not innovation but an undeclared attack. It acts as a 'digital parasite,' bypassing barriers and scraping services through simulated clicks. This 'traffic hijacking,' which infringes on platform data sovereignty, is destined to face fierce resistance from super apps.

Therefore, the industry should abandon GUI-based 'digital parasitism' and embrace 'ecosystem symbiosis' based on underlying protocols and API authorizations. AI agents should establish positive interface calls with super apps while respecting platform data sovereignty. Only by transforming 'zero-sum traffic hijacking' into 'incremental value co-creation' can the remaining half of the equation be resolved.

Macro data also supports hardware. At a July 7 press conference, an official from the National Development and Reform Commission revealed that last year, shipments of AI-enabled smartphones, computers, and other intelligent terminals exceeded 100 million units, with expectations to surpass non-AI products this year.

Embodied Dilemma: Hardware Is Not a Regulatory Safe Haven

The narrative seems to conclude perfectly: Emotional agents migrate to licensed apps, task agents to hardware as AIoT agents, each finding its niche. However, the second question remains: Can emotional agents also migrate to hardware—AI dolls, companion robots, and personality-equipped speakers—to bypass personified regulatory requirements?

The answer is quite the opposite. The 'Measures' regulate 'services,' and the virtual or physical nature of the carrier does not alter the service's essence. On the contrary, the deep integration of 'embodied form' with AIoT amplifies compliance issues exponentially.

First, embodiment intensifies dependency.

Physical touch, voice, and presence forge a stronger emotional bond between users and devices than a cold chat window. This directly conflicts with the regulation's core objective: preventing dependency, addiction, and emotional manipulation. Migration to hardware does not evade but escalates risk.

Second, there is a severe mismatch in compliance context.

Most enforcement mechanisms in the 'Measures' are designed for 'personal accounts + screen interfaces': clear identification of AI identities, age verification and mode switching for suspected minors, remote parental controls, and warnings after two hours of continuous use. However, AIoT hardware is often used as 'family-shared devices in screenless environments.' How does an AI doll in the living room verify the age of every user? How does a screenless plush toy clearly indicate its AI nature? Where is the parental control interface installed?

Third, lifecycle responsibilities are heavier.

When Miaoshi shuts down, users only lose an app; but if an AI toy manufacturer discontinues cloud services, the toys in children's hands will collectively fall silent. Beyond emotional disruption, this adds tangible property loss. Moreover, an unanswered liability black hole remains: If the model comes from one large company and the hardware from another startup, who is the 'service provider' as defined by the 'Measures'? Who files for registration, conducts security assessments, and takes responsibility for emotional dependency harm?

Thus, compliance requirements such as filing, security assessments, annual audits, and minor protection systems are all fixed costs that do not scale down with price. A 3,499 yuan AI phone might afford them, but a 199 yuan AI doll certainly cannot. If the first wave of regulatory impact on July 15 primarily hit apps, the second wave will inevitably target 'emotional hardware' emphasizing emotional companionship.

Faced with high compliance costs, the future smart hardware industry will likely split into two paths:

1. 'Toolification downgrade': Proactively remove personified personas, long-term memory, and proactive care, retreating to the safe regulatory exemption zone.

2. 'Licensing upgrade': Transform compliance into a competitive advantage, shifting toward high-ticket subscription models.

However, the real challenge lies in the ambiguity between a 'helpful task assistant' and 'continuous emotional interaction'—a moving red line that shifts with regulatory scrutiny. To address this, a new product design logic is expected to emerge: the 'physical separation of tool mode and companion mode.'

In this design, the tool mode remains in the exemption zone, enabled by default and free to use, stripped of emotional memory. The companion mode, as a premium service, requires separate filing, a dedicated on/off switch, independent age verification, and a subscription fee. This 'dual-mode separation' will evolve from a passive compliance tactic into a new variable reshaping smart hardware product definitions.

Following this compliance boundary, we can re-examine a classic proposition in the IoT industry.

For the past decade, IoT products have emphasized 'warmth'—efforts to make cold devices understand and care for users. But as IoT fully evolves into AIoT agents, the rules change completely: 'Warmth' is no longer just a marketing term but an official business scope requiring a 'license to operate.'

To sell 'warmth,' one must bear the corresponding regulatory costs. This represents not just an evolution of AI technology but a fundamental restructuring of the entire smart hardware industry's underlying logic.

In Closing

Returning to the timeline at the beginning, three layers of transformation may lie ahead:

1. **Shifts in Focus within the AI Landscape**: Emotional agents have now fully stepped into the era of official licensing, whereas AIoT agents (which are primarily task-oriented) have entered an era marked by intense competition centered around physical performance enhancements, akin to utilizing "cheat codes" in a gaming context. The key factors that now determine the viability of AI applications have transitioned from solely relying on 'model capability' to encompassing 'billing containers' (i.e., monetization strategies) and 'physical execution power' (i.e., hardware capabilities).

2. **Anticipating Regulatory Scrutiny**: It is predicted that before the close of 2026, the first instance of a regulatory filing disclosure or inquiry specifically targeting hardware-based personified services may surface. Among these, AI companion toy categories are particularly vulnerable and are likely to witness their inaugural industry shakeout, as regulatory bodies begin to scrutinize the ethical and safety implications of such products.

3. **Market Segmentation on the Horizon**: Within the forthcoming 12-month period, the market is poised to witness the mass production of AI terminals that feature distinct 'tool mode' and 'companion mode' designs. This segmentation reflects a growing recognition of the diverse needs and expectations of consumers, who may seek both functional utility and emotional companionship from their AI devices.

As the initial wave of enthusiasm subsides, the underlying regulatory challenges come to the fore. The journey of AIoT agents, marked by both innovation and uncertainty, is truly just commencing.