—·
In China’s AI phone market, the contest is shifting from model branding to OS-level execution rights. The winners may be the brands that can turn assistants into operating systems.
A smartphone assistant that can draft a reply is no longer enough. In China’s premium handset market, the more consequential contest is whether an AI agent can read a calendar entry, open the right app, handle a file, trigger a device control, and complete a task across phone, tablet, PC, car, and home screen without breaking trust or platform rules. That is why the most revealing news in early March was not another benchmark claim for a language model. It was Xiaomi’s decision to begin limited testing of Xiaomi miclaw on March 6, 2026, positioning it as a mobile agent product built on its MiMo model and aimed at system-level task execution (finance.sina.com.cn). It came as Huawei was already building out its Harmony Agent Framework and Honor was extending YOYO into a broader agent platform tied to Alibaba’s Qwen models and Qualcomm’s on-device stack (caixinglobal.com) (alibabacloud.com).
The shift matters because China’s smartphone market is large enough to sustain a distinct AI architecture if domestic brands can turn agents into a system function rather than a marketing feature. China shipped 20.696 million smartphones in January 2026, according to the China Academy of Information and Communications Technology, even after a year-on-year decline of 15.6 percent for the month; 32 new smartphone models were launched that month, up 28 percent from a year earlier (gma.caict.ac.cn). Meanwhile, MIIT said in July 2025 that China already had more than 100 AI terminal products on the market, including AI phones, AI PCs, and AI glasses (miit.gov.cn). This is no longer a pilot phase.
The deeper question is not which company has the best model. It is which company can secure the right to execute on the device.
For more than a decade, mobile competition turned on hardware, app stores, and cloud services. The next layer is action orchestration. An AI phone agent becomes strategically valuable when it can move from answering to doing: opening a document, editing a file, reserving a slot, calling a transport service, or carrying context from one device to another. That is why Chinese OEMs are concentrating on OS-level hooks and built-in service access rather than only chatbot interfaces.
Xiaomi’s miclaw is a clear example. The company said on March 6, 2026 that it had started closed testing of a mobile AI agent built on its in-house MiMo large model (finance.sina.com.cn). Coverage from CGTN described the system-level agent as having more than 50 capabilities, including reading and writing text messages and files, controlling smart-home devices, and operating built-in system tools on smartphones (news.cgtn.com). That list is more important than the branding. It signals that Xiaomi wants the assistant to sit inside HyperOS as an execution broker.
Huawei is pushing the same direction from a different architectural starting point. At HDC 2025 in Dongguan on June 20, Huawei launched HarmonyOS 6 and introduced the Harmony Agent Framework, with more than 50 Harmony-powered agents slated to go live, according to event coverage summarizing the official announcement (dongguantoday.com). Huawei’s advantage is structural: HarmonyOS is not merely a skin over Android in the China market. It is the company’s own operating environment, which gives it stronger claims over service invocation, identity, device coordination, and UI rules than Android OEMs usually enjoy.
Honor’s position is more hybrid, but that may turn into a strength. At MWC 2025, Honor demonstrated a GUI-based mobile AI agent that could read the screen and book a restaurant via a third-party service while checking calendar and traffic information (wired.com) (prnewswire.com). By March 2025, Alibaba Cloud said Qwen Max and Qwen-VL had been integrated with Honor’s YOYO assistant (alibabacloud.com). In other words, Honor is building its agent layer through alliance and integration, not by owning the entire OS stack.
That difference points to the core market split in China’s AI phone race. Huawei seeks vertical command. Xiaomi seeks system reach inside a broad device estate. Honor seeks agent fluency across a more modular partner ecosystem. All three are competing for the same layer: the software authority to carry user intent into action.
Xiaomi’s old strength was hardware volume plus a sprawling AIoT network. The AI-phone moment gives the company a chance to turn that scale into agent utility. A conventional assistant can answer a question about a flight. A system-level agent can notice the ticket, check commute times, create reminders, set the alarm, and coordinate with connected devices. CGTN’s description of miclaw’s flight-ticket use case is revealing because it places the agent between notification, calendar, weather, and alarms, which is exactly where execution value sits (news.cgtn.com).
This matters because Xiaomi is not trying to win by becoming another general-purpose model vendor. It is trying to make HyperOS harder to leave. In a smartphone market where hardware cycles are lengthening, the retention value of a good agent may come from accumulated permissions, user habits, smart-home bindings, file history, and device-to-device continuity. The assistant becomes a switching-cost machine.
The company also has a scale base from which to attempt that move. Canalys said AI-capable smartphones already accounted for 22 percent of mainland China shipments in 2024 (canalys.com). Counterpoint reported that Xiaomi held 16.6 percent of China smartphone sales in the first quarter of 2025, while Huawei held 19.5 percent (counterpointresearch.com). In such a market, the first useful agent layer could reshape premium positioning faster than another camera upgrade.
A second reason Xiaomi matters here is timing. miclaw arrived not as a concept tease but as a controlled test in March 2026, after the Chinese market had already spent a year being conditioned to the language of AI phones. MIIT’s mid-2025 statement that China had already surpassed 100 AI terminal products shows that OEMs are moving in a policy environment where “AI terminal” is an accepted industrial category, not a speculative one (miit.gov.cn). Xiaomi is trying to ensure HyperOS becomes one of the places where that category finds its practical form.
Huawei’s AI phone strategy is harder to copy because it sits on top of an increasingly distinct OS and ecosystem. Counterpoint data reported in October 2025 showed HarmonyOS at 17 percent of China’s smartphone operating-system market in the second quarter of 2025, ahead of iOS at 16 percent for the sixth consecutive quarter (tech.yahoo.com). Earlier Counterpoint figures for the fourth quarter of 2024 put HarmonyOS at 19 percent in China, ahead of iOS at 17 percent (gizmochina.com). Those numbers do not make HarmonyOS the dominant mobile OS, but they do make it a serious domestic execution environment.
That distinction matters for agent phones because an assistant is only as capable as the permissions it can safely hold. Huawei’s Harmony Agent Framework gives the company a way to formalize agent access as part of the operating environment rather than bolt it on as an app privilege. Coverage of HDC 2025 reported that Huawei paired HarmonyOS 6 with HMAF and planned 50-plus intelligent agents across work and personal scenarios (dongguantoday.com) (gizmochina.com). This is the clearest case in China of an OEM trying to define an agent-native OS stack.
The commercial challenge is ecosystem depth. Huawei can build execution rights more cleanly than Android-based rivals, but it still needs service availability and developer participation. At HDC 2025, Huawei said the conference would draw more than 12,000 developers, reflecting the scale of its ecosystem push (dongguantoday.com). Yet even favorable reporting noted that Huawei’s app ecosystem remained short of its own targets (gizmochina.com). An agent can only be as useful as the tools and services it can touch.
Still, Huawei has already shown one of the strongest real-world examples of ecosystem migration. The Pura X, announced on March 20, 2025, was reported as the first phone to launch with full pre-installation of HarmonyOS 5 in China, marking a more complete move away from Android compatibility on that device line (en.wikipedia.org). I am using this as a timeline marker rather than as a primary source on features, because the stronger point is supported by Huawei’s own 2024 annual-report release stating that Huawei devices were “back in the fast lane,” with major HarmonyOS ecosystem progress (huawei.com). The significance is not brand theater. It is that Huawei can now tie the AI assistant, the operating system, and the device family into one control plane.
Honor’s strategy looks less self-sufficient than Huawei’s, but it may be more adaptable in the short run. The company’s MWC 2025 demonstration of a UI agent booking a restaurant through a third-party service was one of the first visible examples of a phone maker treating the assistant as a doer rather than a talker (wired.com). The setup also mattered technically. Honor said it worked with Qualcomm to keep personal data on device and build a personal knowledge base over time (wired.com). That indicates a mixed architecture: on-device personalization plus cloud-scale model support where needed.
Honor then reinforced that model partnership route. Alibaba Cloud said in March 2025 that Qwen Max and Qwen-VL were integrated into Honor’s YOYO assistant, and that the tie-up was part of a broader push to bring generative AI to edge devices (alibabacloud.com). Another report on the collaboration said Alibaba-based agents such as AutoNavi and Fliggy Travel were available on an Honor AI phone, pointing to a strategy built around high-value service entry points rather than total OS independence (asiatechwire.com).
This gives Honor a different kind of leverage. Instead of owning the base OS in the Huawei style, it can try to own the orchestration experience: the layer that interprets intent, reads screen state, and moves between services. In that sense, the Honor agent platform is best understood as an execution fabric resting on partnerships with model providers, silicon vendors, and service ecosystems.
Honor has also been explicit about broadening the concept beyond a single phone interface. Caixin reported that the company launched “Honor Lobster Universe,” a platform meant to support AI agents across PCs, tablets, and smartphones (caixinglobal.com). Whether that branding lasts is less important than the direction: the company is arguing that the winning agent is not confined to the handset screen. It must carry state and authority across devices.
China’s AI phone race is unfolding inside a regulatory architecture that now affects distribution, feature design, and vendor credibility, not merely legal compliance. In January 2026, the Cyberspace Administration of China said 446 new generative AI services completed national filing in 2025, while 330 application functions or services that directly call filed models completed local registration, for 776 newly recorded services and functions in 2025 (cac.gov.cn). A separate CAC-related report said that by the end of 2025, 748 generative AI services had completed filing cumulatively (cac.gov.cn). The discrepancy is not trivial bookkeeping; it reflects the fact that China is regulating at multiple layers—base models, downstream functions, and local deployments. For phone makers, that means the AI agent is not one product from a compliance standpoint. It is a chain of regulated components: model access, device-side invocation, service integration, and output labeling.
That has direct consequences for how Xiaomi, Huawei, and Honor can commercialize system-level agents. A phone assistant that only answers prompts resembles a conventional AI service. A phone assistant that can open messages, summarize documents, call maps, book travel, and manipulate system settings starts to look more like a privileged digital operator. In November 2025, CAC announced enforcement actions against mobile internet applications that violated rules on labeling AI-generated and synthetic content, citing failures to provide explicit labels, embedded metadata, and user declaration functions (cac.gov.cn). Applied to AI phones, that suggests at least three likely product requirements: first, visible disclosure when an agent generates or modifies content; second, machine-readable logging or provenance for downstream sharing; and third, tighter permission design around cross-app actions. In practice, the OEM that builds these controls cleanly into the OS will have an easier time scaling agent features than one that treats them as post-launch patches.
The demand side reinforces that point. CAC reported that China’s generative AI user base reached 515 million by June 2025, up 266 million from December 2024 (cac.gov.cn). At that scale, governance stops being a niche policy issue and becomes a mainstream consumer expectation. Users will not only ask whether an agent is useful; they will ask whether it misfires, whether it leaves records, whether it fabricates content into chats or documents, and whether they can tell what happened after the fact. For premium handset brands, trust therefore becomes a measurable commercial variable. The vendor that can show clean permission prompts, visible action trails, and reliable content labeling will have an adoption advantage over one offering a flashier but more opaque assistant.
That is the hidden reason the execution layer matters so much. The closer the agent sits to the OS, the more it can do. The more it can do, the more it enters regulated territory involving identity, records, content provenance, and data protection. In China’s market, that does not merely raise the cost of building an agent; it changes which firms are best positioned to win. Huawei’s tighter OS control, Xiaomi’s embedded HyperOS reach, and Honor’s partner-dependent architecture each imply different compliance burdens. The engineering problem and the compliance problem are becoming the same problem.
It would be easy to frame this race as China versus Apple or Android. That misses the more important development. What is emerging is a third logic of mobile competition in which the key strategic asset is not the app store alone, nor the handset spec sheet alone, but the agent layer that decides how intent becomes action. Xiaomi is trying to build that layer inside a broad Android-derived and AIoT-rich environment. Huawei is using HarmonyOS to make the agent a native component of a domestic operating system. Honor is assembling the layer through alliances with model, silicon, and service partners. These are not cosmetic variations on the same strategy. They are three distinct answers to the question of who controls execution rights on the device.
The local market is large enough for that experiment to matter beyond China. CAICT’s January 2026 figures show that even in a soft month, China still shipped 22.866 million mobile phones, of which 86.9 percent were 5G models (gma.caict.ac.cn). IDC data reported in January 2026 said Huawei shipped 46.7 million smartphones in China in 2025, taking a 16.4 percent market share for the year (chinadaily.com.cn). Canalys, meanwhile, said AI-capable smartphones had already reached 22 percent of mainland shipments in 2024 (canalys.com). Put differently, tens of millions of devices are already entering the market with the hardware headroom and commercial framing needed for on-device or hybrid AI features. That is enough scale to influence developer roadmaps, shape service partnerships, and create user habits around agent-led interaction before similar norms fully settle elsewhere.
There is also a second-order effect that the industry may be underestimating. Once an agent becomes the default route into tasks, the smartphone brand gains bargaining power over apps and services. If the agent decides which map service to call, which travel workflow to trigger, which productivity tool to open, and which device to wake up, then the home screen loses strategic centrality. Discovery shifts upward from icons to orchestration. In that world, being the default callable service matters more than being the most downloaded app, and being deeply integrated with the OEM matters more than merely having a strong standalone brand. That could reorder value across China’s digital ecosystem, benefiting services such as AutoNavi, Fliggy, or Xiaomi’s own device network if they become preferred execution endpoints inside handset-native agents.
That is why the next twelve months matter. By the second half of 2026, China’s leading phone makers should move from controlled demos and limited tests toward clearer commercial agent tiers across flagship lines. The crucial indicators to watch are not benchmark scores but operational ones: how many third-party services each agent can reliably invoke, how often a multistep task completes without human correction, what kinds of permissions are exposed to users, and whether these features remain confined to premium launches or move down into broader volume bands. The company that wins will not necessarily have the smartest model. It will be the one that proves three things at once: reliable task completion, acceptable governance, and enough ecosystem access to make the agent feel native rather than theatrical.
A concrete policy recommendation follows from that. The Cyberspace Administration of China and MIIT should publish a more specific compliance framework for consumer-device AI agents by the fourth quarter of 2026, covering on-device permissions, action logging, synthetic-content labeling, and cross-app invocation standards. China already has filing and labeling systems for generative AI services, but smartphone agents need more explicit rules because they sit inside daily operating systems, not only cloud interfaces (cac.gov.cn) (cac.gov.cn). Clearer standards would reduce uncertainty for OEMs and developers alike.
The commercial forecast is equally specific. If current integration work continues, by Q4 2026 the premium end of China’s smartphone market will be judged less by whose model can answer a prompt and more by whose phone can complete a chain of actions across files, apps, and companion devices with the fewest taps. That would mark a genuine shift in mobile competition. The smartphone would no longer be defined mainly by the apps it hosts. It would be defined by the agent layer that decides how those apps get used.
Honor’s MWC 2026 Robot Phone made a larger point: in China’s AI phone race, the next contest is over hardware that can perceive, move, and act with the OS.
Xiaomi and Huawei are pushing “AI agent” phone assistants that can invoke tools and act. The trade-off is governance: what’s allowed, what’s logged, and who can audit it.
China’s agent-phone wave is moving from demos to end-to-end task execution, forcing handset makers to harden tool permissions, user confirmation, and compliance-grade logging.