Smart Agent AI in Phones is Rising with Dimensity

< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">

2026-05-15 15:36:18 Source: QbitAI

MediaTek is Becoming the Full-Stack Enabler for the Implementation of Agentized Experiences

##### By Klexi from Aofeis Temple

QbitAI | Official Account QbitAI

Agentization is becoming the most intensively explosive direction in the AI industry.

In developer communities, agent frameworks like OpenClaw and Hermes, which feature cross-session persistent memory and self-learning capabilities, have expanded at an astonishing rate since their emergence this year.

Hermes Agent gained 110,000 GitHub stars within ten weeks and just surpassed OpenClaw in May to become the global daily active leader on OpenRouter.

Data also confirms this wave. The volume of autonomous agent tasks grew from 120 million per day in 2025 to 870 million per day in 2026, achieving a sevenfold increase in one year.

The application side shows a similar trend. Whether on the App Store, Google Play, or third-party app stores, agent AI applications have experienced a super surge of up to four times.

Agents have already become efficient, all-weather productivity tools for everyone.

This wave is penetrating every terminal. The real battleground has shifted to where users can reach every day.

However, truly implementing the experience is never something a single chip can solve alone.

How to balance computing power and power consumption, how to transform the system from passive response to active perception, and how to truly integrate the fragmented application ecosystem.

These three layers of issues determine whether the agentized experience can move from concept to everyone's daily life.

Amidst this surging wave, MediaTek presented its judgment at the latest Dimensity Developer Conference—

By leveraging a full-scenario chip matrix, development toolchain, and cloud AI acceleration technology, it aims to become the full-stack enabler for implementing agentized experiences.

Agentization is Redefining Terminal Experiences

The explosion of OpenClaw marks the formation of a new AI paradigm.

Past AI interactions were passive, with users and models engaging in Q&A, each dialogue window independent of the others.

What frameworks like OpenClaw's Harness truly achieve is giving agents cross-session persistent memory and autonomous execution capabilities, turning AI into a partner that can get work done.

The impact of this evolution has already spread from developer communities to the daily lives of every ordinary user.

Over the past year, the volume of autonomous agent tasks has seen a sevenfold increase, from 120 million to 870 million per day, driven by real user behavior.

Simultaneously, the competitive dimensions of AI experience are undergoing a fundamental shift.

Model parameters and chip computing power are no longer the decisive factors for AI experience. Integrating cross-application, cross-terminal, and full-scenario user experiences has begun to become the necessary path for agentization.

This shift in competitive dimension directly points to a core question: Where should this agentized experience be implemented?

The answer can only be: on the device side.

For agents to be truly valuable, they need to continuously perceive user context, understand usage habits, and be available for invocation at any moment the user needs.

Device-side equipment, carried on-person, always online, and accumulating the user's most complete behavioral data and application ecosystem, is naturally the best carrier for agentized experiences.

However, the challenges faced in implementing on the device side are far more complex than imagined.

As more and more applications begin to utilize the NPU, issues of different applications competing for NPU resources have already started to appear.

A deeper contradiction lies in the fact that AI functions on today's smartphones are highly fragmented.

The fundamental reason behind this is that the team structures and functional setups of smartphone manufacturers are already well-established. AI has been superimposed onto this highly mature system, creating a natural distance from the native system's underlying layers.

This structural dilemma cannot be solved by a single technological breakthrough.

It requires someone at the very bottom of the entire chain to simultaneously integrate computing power, systems, toolchains, and the application ecosystem—and this is precisely the natural position for a chip manufacturer.

This is exactly where MediaTek is positioning itself.

At this year's MediaTek Dimensity Developer Conference (MDDC), a new industry consensus is being repeatedly mentioned: The AI war is, in essence, a war of ecosystems.

The key to victory is enabling developers to build agent applications more efficiently, thereby allowing users to perceive seamless agentized experiences in more scenarios.

To this end, MediaTek has chosen to step out of the pure chip performance competition and切入 as a full-stack enabler. Through a full-stack layout encompassing a full-scenario chip matrix, system engines, development toolchains, and cloud AI acceleration technology, and in collaboration with the Dimensity ecosystem, it aims to define a seamlessly flowing, omnipresent agentized experience.

Enabling Seamless Flow of Intelligent Experiences

For agentized experiences to be truly implemented, developers face three layers of overlapping real-world challenges.

• At the terminal layer, agents require powerful full-modal computing power while also needing low-power, always-on perception. There is a fundamental contradiction between these two, which no previous chip architecture could fully satisfy.
• At the system layer, the AgentOS needs to possess truly proactive service capabilities while also ensuring user privacy and security. This is a dual proposition that must be solved simultaneously.
• At the application ecosystem layer, the fragmented application ecosystem needs to be integrated and coordinated, enabling the agentized experience to be truly cross-device, cross-scenario, and omnipresent.

The Dimensity mobile platform offers a complete solution to these three layers of challenges, spanning from chips to systems to developer toolchains.

The core contradiction at the terminal layer is power consumption.

Making the phone always "listen attentively" means the perception module must run 24/7.

The dual-NPU architecture of the Dimensity 9500 was born precisely for this purpose. The ultra-performance NPU handles heavy computation for complex tasks, while the ultra-efficiency NPU (eNPU) is dedicated to lightweight Always-On perception.

In this mode, the power consumption of resident lightweight AI models is reduced by 42%. Phones truly gain the capability for continuous environmental perception, covering vision, hearing, and location across all modalities. This provides a real computational foundation for what agents can actively perceive and do.

This dual-NPU architecture changes how on-device AI works, making Always-On perception a常态 capability, freeing developers and users from having to make difficult trade-offs between battery life and experience.

With this foundation in place, the evolution at the system layer gains meaning.

The Dimensity AI Agent Engine 2.0 introduces Dimensity SensingClaw technology, endowing the AgentOS with full-time perception capabilities for vision, hearing, and location.

This framework can automatically decompose complex tasks based on the current scenario and accomplish them through cross-application collaboration.

For example, if you're waiting in line for a number call, SensingClaw can help track the announcement audio and proactively push a reminder when your number is called.

Or, if you want to buy a new phone, SensingClaw can help you compare prices across different shopping platforms.

What SensingClaw truly changes is the way agents介入 users' lives.

The capability for active perception and autonomous execution means the agentized experience gains true proactivity.

This is a key step for on-device AI transitioning from a tool属性 to a service属性.

Furthermore, MediaTek has collaborated with smartphone manufacturers like OPPO, Xiaomi, and Transsion to launch the system-native Claw, sinking the capabilities of active perception, autonomous execution, and seamless cross-device flow to the system layer of phone manufacturers. The Agent transforms from a tool into the system itself.

No matter how strong the underlying capabilities are, they ultimately need to be perceived by users through applications.

Developers are the key variable determining whether agentized experiences can be implemented at scale and the gatekeepers for the final mile of the entire chain.

However, the development门槛 for agentized applications is much higher than for ordinary apps—model deployment processes are cumbersome, on-device memory constraints are strict, Always-On perception capabilities are difficult to access, and the learning cost from zero to implementation remains high.

These four barriers叠加 together cause many developers to hesitate before even starting.

The newly released Dimensity AI Development Kit 3.0 at this conference addresses these real development pain points with four upgrades—

• Visual deployment tools improve efficiency by 50%;
• The LowBit compression toolkit improves compression rate by 58% under同等 quality, directly lowering the on-device memory门槛;
• The eNPU development toolkit opens up the Always-On perception capability of the ultra-efficiency NPU to developers for the first time;
• Dimensity AI Partner compresses the end-to-end deployment work that used to take 5 days down to half a day.

These four upgrades point towards the same goal: allowing developers to shift their focus from磨合 tools and processes to creating experiences.

The more user-friendly the toolchain, the faster the ecosystem expands.

Over the past two years, downloads of the Dimensity AI Development Kit have accumulated an increase of 440%, and生态伙伴 growth has reached 240%.

Behind these numbers reflects a self-reinforcing momentum正在 forming within the entire Dimensity ecosystem—

The more developers, the richer the applications, the better the experience, attracting more developers to join. This flywheel has already started turning.

Ecosystem Power, the Ceiling of Experience Innovation

Having powerful chips, intelligent systems, and complete toolchains together only grants an entry ticket in the process of scaling agentized experiences.

The true上限 is determined by ecosystem power.

The AI industry has fully entered the era of "ecosystem-defining experience." The explosive growth speed of agent AI applications far exceeds the消化 capacity of any single manufacturer. No single company can independently complete the full-chain construction from computing power, systems, toolchains to the application ecosystem.

The competitive dimension itself is shifting. Players who can establish standards across the entire chain, gather developers, and connect upstream and downstream truly possess the qualifications to compete in the agentization era.

An ecosystem is a chain, with different members at each link. The key is to work backwards from the end-user experience, identify one's position within the chain, and clarify which key links to focus on supporting.

The AI war is, in essence, a war of ecosystems. What truly decides victory is the ability to enable developers to build agent applications more efficiently and allow users to perceive seamless agentized experiences in more scenarios.

The one capable of achieving this is the enabler positioned at the very bottom of the entire chain, simultaneously connecting computing power, systems, toolchains, and the application ecosystem. It determines how fast and how far the upper-layer experiences can run.

The role boundary of chip manufacturers is being重新 defined by this era, transitioning from suppliers providing computing power to infrastructure providers defining experience standards.

In the first two MDDCs, MediaTek brought new chips. This year, it did not. However, the information density of this year's conference is no less than before.

This signal reflects that the role MediaTek has chosen to play next is precisely that of the重新 defined "enabler" for chip manufacturers.

What remains unchanged is that at历届 MDDCs, MediaTek has consistently坚持 positioned Agentic AI as a core议题.

This consistency itself is a manifestation of judgment.

When the industry hadn't yet formed a consensus, betting on a direction and delivering on it every year.

The five characteristics of Agentic AI UX defined last year have now all seen concrete technological and product implementations.

From the perception capabilities at the chip底层, to proactive execution at the system layer, and再到 the developer tools at the application layer, the implementation at each layer validates that the direction of this judgment is correct.

This布局 was确立 from the very beginning of chip design. MediaTek was the one that built the road at the底层 before agentized AI became an industry consensus.

From the dual-NPU architecture of the Dimensity 9500, to the Dimensity AI Agent Engine 2.0, and再到 the Dimensity AI Development Kit 3.0, every layer converges towards the same direction—

Turning the agentized experience into a result that is perceivable by users, implementable by developers, and sustainably evolvable by the industry.

This logic is precisely MediaTek's concrete interpretation of the role of "full-stack enabler." They are becoming the AI infrastructure provider for implementing agentized experiences through a full-scenario chip matrix, system engines, development toolchains, and AI acceleration technology.

MediaTek is turning this capability into infrastructure that anyone can build upon and deliver on the Dimensity platform.

The boundaries of the Dimensity ecosystem are becoming the boundaries of the agentized experience.

_All rights reserved. Unauthorized reproduction or use in any form is prohibited._