Speaking at the main forum of Alibaba's 2026 Apsara Conference, Fang Fei, president of Honor's product line, laid out the company's technology roadmap for on-device artificial intelligence. Honor will partner with Alibaba to build five vertical-domain models and solutions, which will debut in the forthcoming Honor Magic9 series as the company tries to move phone agents from merely "understanding" tasks to actually carrying them out.

Phone AI Competition Shifts From Model Capability to System Coordination

As AI moves faster into consumer electronics, the terms of competition in smartphone AI are changing.

Industry analysis from IDC suggests on-device AI is no longer judged solely on model parameter counts or single-benchmark scores, but increasingly on how well perception, memory, reasoning, planning and execution work together. Put differently, a phone's AI has to do more than parse a command: it needs to read its surroundings, remember user habits, and marshal the operating system and apps to finish complex tasks.

Against that backdrop, Honor plans to upgrade MagicOS into what it calls a companion-style multimodal agent operating system, or Agentic OS, rebuilding six layers of technology spanning hardware, system and ecosystem. The goal is to shift phone interaction from user-initiated operation to intent-driven, proactive service.

Fang said devices will no longer serve only as containers for running apps, but as platforms where various agents operate, and that the operating system is the key to making that shift.

MagicOS 11 Adds a System-Level Agent Harness

Once a large model has interpreted an instruction and reasoned through a task, the system still has to sense the environment, break the task down, call the right tools and execute. To address agents that can understand but struggle to act, Honor has introduced a system-level Agent Harness architecture in MagicOS 11.

The architecture ties together the model, the system environment, task planning, tool invocation and execution, letting an agent decompose a user instruction into sequential steps and adjust on the fly as conditions change.

Say a user asks for something that spans several apps: the system has to recognize the current screen state, call the relevant apps and services, fill in information, navigate between pages and confirm the result. Agent Harness translates the model's reasoning into an operational flow the system can execute reliably.

Honor and Alibaba Jointly Build Five Vertical-Domain Models

On the model side, Honor and Alibaba are pursuing joint device-cloud development. Using Honor's MagicOS and Alibaba's Qwen Intelligence as the technical foundation, Alibaba supplies the base large-model capability while Honor contributes device system technology and real usage scenarios, with the two co-developing five vertical-domain models and solutions tuned for phones.

Rather than simply plugging a general-purpose large model into a handset, the partnership adapts models to specific scenarios including productivity, coding, image creation, fitness and health, education and travel, aiming to sharpen how well models understand, invoke and execute in a real device environment.

According to Fang, the jointly developed models and system framework helped Honor's new YOYO assistant score 91.8% composite task accuracy in testing, with automated operations taking 3.6 seconds on average, support for chains of more than 100 consecutive steps, and a 90% end-to-end task completion rate. Fang said those figures improve on the previous version.

The numbers reflect capability measured in Honor's internal or partner testing; real-world performance will still hinge on task type, how well apps are adapted, network conditions and system permissions.

The Phone as a Hub for Multi-Device AI Coordination

Honor argues that devices act as the connective tissue between user scenarios, model computing power and services. In its view, the phone will stop being a standalone piece of hardware and instead become the scheduling hub linking phones, tablets, watches, PCs and other smart devices.

Beyond the handset itself, distributed devices can feed in richer environmental information. Smart glasses can capture what a user sees, a watch can supply fitness and health data, and newer form factors such as Robot Phone attempt to extend AI beyond the phone screen into the physical world.

The Honor Magic9 series is set to launch on September 28, carrying the agent models and full solution stack developed with Alibaba, along with upgraded mobile imaging.

Fang said Honor will keep working with Alibaba, supply chain partners and developers to build out its on-device agent ecosystem, extending agent capabilities to more devices and use cases so that AI moves beyond chat and Q&A into everyday tasks for ordinary users.