The 'Pseudo-Native' Dilemma of AI PCs: Examining the Real Barriers to Agent Implementation Through the Lens of Doubao Phones

[Ebrun Original]

Last month, NVIDIA CEO Jensen Huang prominently promoted the 'AI PC with Agent' at Computex, describing it as 'the biggest PC revolution in 40 years.' Intel quickly followed suit, proposing the concept of the 'Agent PC,' vowing to make it a 'new gateway' for the AI era that is 'accessible to all, cost-effective, secure and reliable, and rich in scenarios.' However, peeling back the marketing veneer reveals a significant gap between current AI PCs and true 'AI Agent-native' systems—a gap that becomes particularly glaring when compared to the 'native agent' path represented by Doubao phones.

I. AI PC: Hardware Upgrade-Driven 'Pseudo-Native'

Several industry analysts have been blunt in pouring cold water on the concept of the 'AI PC with Agent.' Leonard Lee, Principal Analyst at neXt Curve, pointed out: 'AI PC with Agent is a strange term. Depending on the use case, the last two generations of PCs have already had the capability to run agents.' Jim McGregor, Principal Analyst at Tirias Research, also believes that while AI PCs equipped with agents have improved capabilities, the related application ecosystem remains scarce.

In essence, the so-called 'AI PC with Agent' is merely an evolutionary continuation of early AI PCs—enhancing AI capabilities by embedding software forms like PC managers and agents onto the traditional x86/Windows architecture. While NVIDIA's RTX Spark PC is touted as 'the world's first Windows PC superchip built for personal agents,' its Arm-based N1X chip may have compatibility issues with existing x86 applications, limiting the significance of large-scale upgrades in the short term.

Intel's 'Agent PC' is even more straightforward—it is essentially an 'Advanced AI PC,' an AI PC possessing 'full-chain execution capabilities including intent understanding, task decomposition, tool invocation, iterative execution, and memory retention.' This definition precisely exposes the problem: The 'agentization' of AI PCs involves layering capabilities onto existing hardware architectures, rather than redesigning the system from the ground up for Agents.

II. Doubao Phone: From 'Feature Stacking' to 'Native Agent'

In stark contrast to the AI PC's path of 'hardware upgrade and software stacking,' the Doubao phone represents a paradigm that rethinks Agent interaction from the operating system's foundation. Ni Fei, Senior Vice President of ZTE, explicitly stated that the limitation of current AI phones is that 'explorations like intelligent conversation and image creation at this stage feel more like stacking AI features onto phones. In users' daily operations, they still need to jump between multiple apps and click repeatedly.'

The next phase for AI phones will shift from 'feature stacking' to 'native agent,' completing the leap from 'users being required to adapt to the phone' to 'the phone actively understanding needs and autonomously completing tasks.' The first-generation Doubao phone (Nubia M153) has already demonstrated the power of this 'native' approach: Users can make it operate across applications through natural language—comparing prices, editing photos, checking tickets, placing orders, sending messages. The AI acts 'like a real person holding a phone, step by step opening the app, identifying the interface, and completing the task.' It relies on the Android system-level 'INJECT_EVENTS (simulated touch)' underlying permission to support cross-application voice execution. This system-level permission-granted cross-application operation capability is unattainable by any current 'AI PC software-stacking solution.'

The second-generation Doubao phone goes further. Ni Fei proposed that the 'native agent' stage should possess four core benchmarks: natural interaction, task execution, habit memory, and underlying security. Hardware design is also being reconsidered around the Agent—no longer centered on apps, but reserving system resources for continuously running Agent tasks.

III. Core Gap: System Design Constraints vs. Hardware Spec Stacking

The fundamental gap between AI PCs and Doubao phones lies not in computational power, but in system design constraints.

First, permissions and closed-loop capabilities. AI phones (like the Doubao AI phone), despite having NPU computing power of only 40-100 TOPS, possess cross-application data invocation permissions. They can automatically compare prices, book tickets, and sync schedules—these high-frequency tasks can be completed without an internet connection. In contrast, the so-called 'agents' on the PC side mostly remain confined to single-step Q&A within dialog boxes, unable to penetrate application boundaries to complete closed-loop tasks.

Second, the depth of coupling between hardware and the system. The Doubao phone makes way for the Agent from chip selection to system permissions—hardware is re-evaluated around the Agent, rather than 'pasting an AI layer' onto an existing architecture. AI PCs are the opposite. Whether it's NVIDIA's RTX Spark or Intel's Core Ultra, the core narrative remains 'higher computing power, larger memory, stronger models,' not 're-architecting the operating system for the Agent.'

Conclusion

The current 'agent' narrative of AI PCs more closely resembles a marketing breakout by PC manufacturers against the backdrop of an overall sluggish market (IDC estimates a 11.3% decline in global PC shipments by 2026). It has indeed made breakthroughs in hardware computing power—the Core Ultra 358H can reach up to 180 TOPS, capable of running 35B MoE models smoothly—but these represent 'quantitative' accumulation, not a 'qualitative' leap.

True AI Agent-native is not about 'running large models on a PC,' but about designing interaction paradigms, permission models, and task closure from the operating system's foundation for the Agent. The Doubao phone has demonstrated the feasibility of this path with its system-level permissions and cross-application operations. In contrast, the 'agent' of AI PCs remains stuck in the old paradigm of 'faster hardware, stronger models.' As analysts have said, the so-called 'AI PC with Agent' may indeed be 'overstated.'