Say to your phone, "I need to attend an offline meeting at the convention center this afternoon. I need to arrive 20 minutes early. Please check when I need to leave, book a car, and remind me." What happens next is quick: Xiaoyi checks the meeting time and location in her schedule, uses a map to check real-time traffic conditions, calculates the departure time, books a car, and sets an alarm.
One task, multiple services from three applications, can be scheduled with just one line of code.
—This is a typical scenario of using Xiaoyi (Xiaoyi's AI assistant) showcased at the Huawei Developer Conference HDC 2026.
On the surface, it appears to be an upgrade to voice interaction, but the underlying changes are significant: this task spans multiple independent applications such as calendar, map, and clock. In the past, users had to manually open each of them and flip through pages one by one to find the specific service; but now Xiaoyi has connected the entire process, and users only need to say what they want.
Behind this new mobile AI Agent interaction experience lies a comprehensive evolution of HarmonyOS towards an Agent architecture. Based on the newly upgraded HarmonyOS Intelligent Agent Framework 2.0, Xiaoyi is deeply integrated with the HarmonyOS system, becoming the system's intelligent brain and transforming into an intelligent agent that grows within the operating system, evolving towards becoming the most powerful system intelligent agent.
A new era, a new entry point
Over the past two decades, the unified operating logic of the mobile internet has been based on the path of user → find application → find service: different applications provide different services, users actively open the application they need, and then find a specific service within the application.
We've all taken this well-functioning logic for granted, but in reality, the average person has 70-80 apps installed on their phone, and only uses about 10 per day. According to data from a third-party analytics platform, for a typical smartphone user, 62% of apps are never opened even once a month.
These applications aren't bad; they just have narrow use cases and are abandoned after use. Their service value still exists, but new ways of invoking them are needed.
The emerging path of the Agent era is "Intent as a Service": User submits a request → Agent delivers it. Specifically, the Agent judges, schedules, and executes requests, accurately understanding the needs and distributing services. Users don't need to know which Agents/Skills are being invoked behind the scenes; they only need to know that their needs have been met.
In the new era of mobile/AI agents, the nature of the entry point has also changed. In the past, applications competed for the "information entry point," deciding what users saw; in the agent era, the competition is for the "action entry point," deciding what to do on behalf of users.
Agents are the next default entry point, deeper and more frequently accessed than the search box. Whoever controls them controls the distribution power of the next generation of mobile internet. This is why platform vendors in the mobile internet era feel both nervous and excited.
Last month, Google redefined Android as an "operating system" and then as an "intelligence system." Gartner predicts that by the end of this year, 40% of enterprise applications will embed AI agents, compared to less than 5% last year.
Initially, Apple used the App Intents framework to allow developers to expose functional interfaces for Siri, but this was limited by Siri's own capabilities. However, at last week's WWDC, Apple delivered a new solution: Siri now finally has its own independent app form, and its ability to understand user intent and operate the phone has been greatly improved. It's clear that Apple is using a series of new Apple Intelligence features and Siri AI capabilities to attract ecosystem developers.
Google is even more aggressive: Gemini Intelligence is directly embedded in Android 17 as a system-level agent, and the AppFunctions API allows applications to expose functional interfaces, with the GUI serving as a fallback when no API is available. However, currently only twenty or thirty applications are integrated.
Domestically, we are seeing the most cutting-edge developments, unsurprisingly, happening to Huawei.
HarmonyOS's "meta-service" architecture naturally breaks down services into lightweight, installation-free, and cross-device callable modules, and Agent Wave just happens to provide ready-made channels. Huawei has already taken three steps on this path: building the basic capabilities and framework of HarmonyOS AI in 2024, releasing HMAF in 2025 to promote application-level Agentization, and this year's goal is to Agentize the entire operating system.
The three things announced at HDC 2026 are the specific path to this third step.
The first thing: make the system fully skill-based, turning the system into Xiaoyi's senses and limbs.
To truly be effective, Xiaoyi first needs to leverage more system capabilities. Sensing, data, and device control were previously locked at the system's lower levels. Huawei's approach is straightforward: fully digitize them and delegate them to Xiaoyi for unified scheduling.
The figures released at HDC 2026 show that the system applications are fully Skill-based, with over 2,100 system capabilities accessible to Xiaoyi and over 200 system-level data points integrated into Xiaoyi's context. This includes the capabilities of system applications such as Notes, Calendar, Gallery, Bluetooth, Contacts, and File Management. Therefore, with user authorization, Xiaoyi can access the camera, understand screen content, perceive scene states, and coordinate multi-device collaboration. The entire HarmonyOS system capability becomes a readily available toolbox for Xiaoyi, and users no longer need to tell Xiaoyi "open this first, then that."
There are two scenarios that are very suitable for demonstrating the user experience of Xiaoyi as an Agent after it has gained system-level capabilities:
You tell Xiaoyi, "I've signed up for the Chuxiong Half Marathon, can you create a recovery training plan for me?" Xiaoyi's response chain is as follows: First, it breaks down the intention, searches for relevant race information, and determines that the user is in a period of inactivity or recovery, requiring a gentle restart plan; then it accesses system data such as schedule, sports health, and sleep; next, it coordinates the collaborative work of web search, sports health coach, sports health, and other skills; finally, it conducts a comprehensive evaluation, outputs a customized training plan, and writes it to the calendar.
Let's look at cross-device collaboration capabilities: For example, if you're not in front of your HarmonyOS computer but need some critical files, you can tell Xiaoyi, "Pack all the materials related to the Fourier experiment on your computer and send them to my phone, along with a materials list." Even if the computer is closed and in standby mode, or you're unsure of the file names, Xiaoyi can still help you complete the task. The specific steps are: First, Xiaoyi on your phone will connect with Xiaoyi on your PC to find recently saved relevant data on your computer and send you the list for confirmation. After confirming which data we need, Xiaoyi on your PC will use the document management skill to package the files into a compressed file and send it to Xiaoyi on your phone. A cross-device notification will then be sent upon completion.
Moreover, during the use of Xiaoyi, it is transparent to the user which tool it is calling and which step it is executing, making it visible and controllable.
These scenarios involve currently popular industry concepts: A2A and Skill.
The A2A (agent-to-agent) concept was first proposed by Google and IBM last April and subsequently submitted to the Linux Foundation to become a community-driven open-source international standard, currently supported by more than 150 organizations worldwide. Huawei's previously released AI terminal white paper and HarmonyOS intelligent agent framework white paper both emphasized system-level edge/cloud intelligent agent collaboration protocols, making it one of the earliest vendors to implement A2A collaboration as a system-level protocol.
In the traditional AI interaction chain, humans speak to AI, and then AI executes the commands. But the difference with A2A is that agents can directly communicate and collaborate, which requires a smart "manager" or "coordinator".
In the half-marathon scenario just now, Xiaoyi broke down the task and then dispatched multiple sub-Agents and Skills such as "Sports and Health", "Schedule", and "Search", allowing them to negotiate and exchange data on their own through the Harmony Agent Framework (HMAF) protocol, and advance the task in parallel.
In addition to autonomously planning and coordinating various system capabilities or ecosystem partner capabilities and A2A to execute tasks, Xiaoyi can also do something even more interesting:To help users "create tools".
Tell Xiaoyi, "Make a small tool for children to learn English. It supports learning words by taking pictures of objects. After taking a picture and recognizing the object, it displays the English word, phonetic symbol, definition, and pronunciation. The cards can be added to the vocabulary book." By calling the "Code on Flight" text-to-speech capability and combining it with the system's camera and TTS (text-to-speech) capabilities, Xiaoyi can quickly generate an interactive learning tool. No need to download an app or write a single line of code; it's ready to use as soon as you say it.
This goes a step further than having Xiaoyi help you book flights or create training plans: users are no longer just task initiators, but also tool creators.
The second thing: Open up AI capabilities to allow developers to access them flexibly and efficiently.
The system's application of skills and A2A support laid a solid foundation for Xiaoyi. However, if a system-level agent only calls its own services, its limitations will quickly be reached. The second issue to address is: integrating the ecosystem.
At HDC 2026, Xiaoyi Open Platform was also upgraded, opening up multiple access methods: Agent, Skill, MCP, Intent Framework, etc. Third-party applications can choose the most suitable one for their deep integration with Xiaoyi's scheduling system. Currently, more than 500 selected partner Skills are online, and more than 2,000 HarmonyOS intelligent agents are available. For example, Eastmoney's Miaoxiang Skill can already link with the Internet search Skill to complete custom condition stock selection using natural language.
At a deeper level, HarmonyOS has cumulatively opened up more than 20 AI subsystem capabilities, with a daily call volume of 3.68 billion. Capabilities such as information summarization, card recognition, liveness detection, and image super-resolution have been integrated by applications such as JD Finance, China Railway 12306, and Zhihu. Developers can call system-level AI in their applications without building their own models.
To ensure a smooth connection, the HarmonyOS intelligent agent framework has done several specific things:
The upgraded A2A protocol supports direct connection to the device side, allowing leading partners' agents to collaborate locally on the device. Tencent Video has already integrated through this method, allowing users to directly ask Xiaoyi (the AI assistant) to set the playback speed and ask about the plot without leaving the current interface. The A2UI protocol eliminates the need for predefined interfaces for skills; the agent dynamically generates the UI based on the complexity of user commands, and the same skill can be rendered on mobile phones, tablets, and in-vehicle systems. Skills have integrated Huawei account authentication upon listing, and with the help of natural language development tools, it allows developer teams of different sizes and capabilities to quickly integrate.
For users, they only need to state their needs to get results. Xiaoyi can call on various skills/agents to provide users with a more ultimate intelligent experience.
This logic is called "intent as a service". The user states their needs, and after the system understands the intent, it directly matches an Agent or Skill that can complete the task, skipping the step of "opening the App - finding a service".
For example, if three or five friends come over for a meal on the weekend, they might say to Xiaoyi, "Three good buddies are coming today, help me plan a menu." Xiaoyi first generates a menu based on the number of people and their taste preferences, listing the ingredients and quantities needed for each dish. Then, it calls Dingdong Maicai's Agent to add the ingredient list to the shopping cart with one click, and the user can place the order directly after confirmation.
From "What do you want to eat?" to "Ingredients delivered to your door," this single sentence connects the three stages of menu planning, ingredient matching, and fresh food e-commerce. The user didn't open any of the apps.
For developers, the distribution logic has fundamentally changed, and is completely different from the app stores of the past. In the past, it was "users finding apps," with tens of thousands of apps competing for rankings and bidding for recommendation slots.
The logic of "intent as a service" is "users state their needs, and the system finds services". As long as the service is well done and the Agent/Skill adaptation is smooth enough, Xiaoyi will proactively and accurately call the developer's service when matching user needs.
Huawei launched the Tiangong Program last year with an investment of 1 billion yuan, aiming to incubate more than 10,000 AI meta-services, more than 1,000 intent frameworks and MCPs, and more than 5,000 agents—essentially helping developers lower the access threshold.
Among the various access methods, MCP (Model Context Protocol) is an Agent-Tool Communication Standard proposed by Anthropic in 2024, and it is now the de facto industry standard. HarmonyOS's compatibility with MCP is equivalent to opening a door to the external AI ecosystem: no matter which large model ecosystem a developer is in, as long as they comply with this protocol, they can enter Xiaoyi's scheduling system.
Finally, when an agent can schedule services across applications and read system data, security is no longer an additional concern.
At HDC 2026, Huawei released the "HarmonyOS Intelligent Security White Paper," placing security architecture on an equal footing with agent capabilities.
Xiaoyi adopts the HPIC (HarmonyOS Personal Intelligent Computing System) architecture, adhering to the principles of local priority, data minimization, and user control. When cloud computing power is required, AI inference runs in a user-exclusive confidential environment, achieving triple anonymity and blindness of user identity through the OHTTP protocol and end-to-cloud collaborative encryption. System administrators cannot view user plaintext data.
At the application level, when operations such as money transfers and payments are involved, the system will block the automated process and the user must complete it manually; sensitive information such as certificates and work documents are transferred locally on the device, and the cloud only receives the processed results.
HPIC has passed three major security tests conducted by the China Telecommunication Technology Labs (CTTL), meeting the industry's high standards for enhanced-level security. Furthermore, Xiaoyi Claw has become the first terminal manufacturer to receive national-level certification from the China Academy of Information and Communications Technology (CAICT) for its Claw-type intelligent agent.
The third thing: Introducing high-quality large-scale models from the industry, giving Xiaoyi an upgradeable brain.
The third thing concerns the source of Xiaoyi's "intelligence".
Xiaoyi needs to be smart, which requires a smart brain. Interestingly, Huawei decisively chose not to bet on a single model.
The world of large-scale models changes too rapidly; the optimal solution six months ago may no longer be optimal six months later. Huawei's judgment is pragmatic: rather than betting on a specific model, it's better to build a "universal socket" model access architecture, allowing large models like Pangu and DeepSeek to flexibly connect and call upon it. Whoever performs best in a particular scenario, that's who Xiaoyi will use.
It is precisely because of this flexibility that Xiaoyi's user experience becomes increasingly intelligent. Based on the Agentic self-evolving architecture, Xiaoyi can think, plan, and schedule autonomously. The result is that users often don't need to tell Xiaoyi how to do a complex task; Xiaoyi can do it very well on its own.
In practical use, the differences brought about by the flexible integration of models are obvious. When you say to Xiaoyi, "Write me an English business email," Xiaoyi judges this as a multilingual generation task and routes it to a model with stronger English writing capabilities. When you say, "Analyze my exercise data from last month and give me some improvement suggestions," the task involves structured data understanding and personalized reasoning, and the framework may switch to another model that is better at this type of reasoning. Two requests, two paths, but the user only perceives that "Xiaoyi has done it all."
This choice is similar to Apple's multi-model strategy in the Core AI framework, but the level of implementation differs. Apple has opened up the multi-model backend at the Siri level, providing more of a choice and handing it directly to developers, but without much known work on the coordination level.Huawei's work involved building a model scheduling framework at the operating system level, covering the entire chain from intent understanding to task execution, and the switching of models is transparent to the user.
Model scheduling solves the "can it be done?" problem, but there is another equally crucial issue: Xiaoyi doesn't remember who you are.
Just two weeks ago, OpenAI launched Dreaming V3, which completely rewrote ChatGPT's memory architecture: the system automatically synthesizes user profiles across all historical conversations in the background without requiring manual instructions to "remember," and outdated information is automatically updated.
The industry trend is becoming increasingly clear: no matter how strong the model is, if every conversation feels like the first time, the upper limit of the user experience is very low.
During the sharing and exchange at the HDC 2026 breakfast meeting, the person in charge of Xiaoyi introduced Xiaoyi's efforts in "how to remember users".
Xiaoyi has upgraded its memory system: instantaneous memory processes single reasoning, short-term memory covers a single dialogue window, long-term memory accumulates the user's instruction preferences and implicit habits, and full-time memory continuously accumulates accompanying states, ultimately forming a global dynamic profile.
Unlike purely conversational AI, Xiaoyi's memory sources extend beyond chat logs—it includes system-level data such as calendars, health tracking, spending records, and photo galleries. Huawei's positioning is also clear: this memory system, called "HarmonyOS Memory," is a foundational capability of AI OS, not a proprietary feature of the Xiaoyi application. System-level data, multi-scenario devices, and on-device privacy processing—this combination differentiates HarmonyOS at the memory level.
For example, after our annual physical exam, we habitually download the report and save it locally. However, it can be troublesome to compare specific health data from previous years. Once Xiaoyi remembers the data for you, you can ask her, "What is the trend of my triglycerides over the past three years?" Xiaoyi can retrieve the data across different time periods, provide a curve showing the changes, and even offer consultation services in conjunction with a health-related AI.
Memories are merely archives if they don't return to the scene. They only become abilities when they return to the service.
This capability is not limited to a single app, but is cross-application; the system determines when to pop up based on the intent patterns stored in its memory. This "memory storage → return to service" link shifts the "intent as a service" concept from passive response to proactive prediction.
As mentioned earlier, a high-quality agent product requires two things: a good model and a good scaffolding. These two elements complement each other. A good agent framework allows for the maximum utilization of a good agentic model, resulting in a better user experience.For users, the secret to Xiaoyi's ability to make them feel "smart and understanding" and "get things done" lies in an engineering system that allows it to use good models and accumulate more and more memory.
When the operating system becomes the agent
Since its launch in 2008, the App Store has grown into a trillion-dollar distribution and business ecosystem over the past 18 years, following the path of "users downloading apps and apps providing services".
But now, all platform vendors are doing the same thing: migrating the entry point from apps to agents, bringing users a better AI experience and creating new opportunities for developers. The speed of this migration and the speed of ecosystem expansion are likely to be even faster than before. For reference: it took MCP six months from its release to become the de facto industry standard, and A2A took just over a year from its proposal to the participation of over 150 organizations. In the early days of the mobile internet, it took at least three to five years for a development paradigm to go from proposal to widespread adoption.
Huawei's decision to deeply integrate Xiaoyi into the underlying operating system at HDC 2026 is a forward-looking attempt along this new path.
But what's even more thought-provoking is where this path leads? When an agent accumulates your schedule, health data, spending preferences, and work habits, its relationship with you transcends being merely a "tool." It transforms into a collaborator that understands you, requiring no re-learning of who you are, what your preferences are, where you left off last time, or what issues need improvement—it remembers it all.
As for applications, they certainly won't disappear, but new entry points have already emerged.