Ideal vs. HarmonyOS: A High-Stakes Bet on the Divergence of Agent Roadmaps

At the recently concluded Beijing Auto Show, AI remained the dominant theme, yet the center of the competition has quietly shifted. The battleground now appears to have moved from advanced driver-assistance systems to in‑cabin intelligence. As automakers pivot from “safe delivery” in autonomous driving to “enhanced user experience” in the cockpit, mere screen size or entertainment features are no longer sufficient to create a competitive edge. The real contest lies in who can transform the car into an intelligent agent—one that senses, makes decisions, and proactively delivers services.
On the surface, this once again pits the new‑carmakers against the tech giants in the automotive space. The latest expansion of the HarmonyOS Smart Car ecosystem underscores its overwhelming advantages in scale and ecosystem. Meanwhile, among the new‑carmaker contingent, Li Auto has been the most vocal and steadfast in its AI investments, even embedding Agent capabilities directly into the name of its latest flagship, the L9. The rivalry between these two players over the in‑cabin AI agent has become a crucial lens for gauging the industry’s technological trajectory and its commercial prospects.

However, when you peel back the layers of the “intelligent cockpit,” you’ll find that their underlying architectures take diametrically opposed approaches. This isn’t a case of two players on the same track chasing each other; rather, it reflects two distinct technological philosophies offering different answers to the question of “in‑vehicle intelligence.” HarmonyOS’s MoLA 2.0 architecture adheres to the classic “brain + cerebellum” paradigm. Its cloud‑based “brain” (a system‑level agent) functions like a command center, responsible for understanding, breaking down, and orchestrating tasks before delegating specific duties—such as navigation or vehicle control—to domain‑specific “cerebella” for execution. This clear hierarchical structure naturally turns the HarmonyOS cockpit into a service‑distribution platform, making it easy to integrate third‑party capabilities and fostering an expansive ecosystem.

The ideal StreamingClaw architecture, by contrast, takes a more “heavy‑weight” approach. It aims for an end-to-end closed loop of “stream‑level perception–decision–execution.” Various sensor inputs—such as video and audio—flow in real time like a river, demanding immediate processing; otherwise, the information would be lost in an instant. To this end, Li Auto has developed a multi‑agent coordination mechanism centered on “self‑planning and scheduling.” Crucially, its central agent serves not only as an “intent router” but also actively participates in both perception and decision‑making.
The two approaches are neither superior nor inferior; they simply offer different responses to the in-vehicle agent. In terms of design philosophy, the HarmonyOS architecture is built around task distribution, whereas the Ideal architecture centers on physical perception. Put simply, HarmonyOS constructs an efficient “task‑scheduling hub,” while Ideal seeks to forge the entire vehicle into a unified “embodied entity.” Interestingly, the divergent paths these two systems take in intelligent driving technology mirror their respective choices regarding the agent, revealing a profound “path dependency.” In the era when software was considered a “personal creation,” products each bore distinct, individualized traits. Today, large‑scale software development may appear entirely “depersonalized,” yet underlying development styles still follow discernible patterns.
Ideal has extended its end-to-end and VLA (Vision–Language–Action) modeling approach—originally applied to intelligent driving—to Livis, aiming to integrate spatial perception, language understanding, and decision-making into a unified framework, with the ultimate goal of creating a general-purpose agent for the physical world. Reportedly, by 2026, this effort will be subsumed under a unified large-scale model, striving to build a versatile agent tailored for real-world environments. Meanwhile, Huawei’s ADS leverages a world model to construct a cloud‑based training environment, continuing its safety‑oriented architecture that combines “world modeling” with simulation‑driven optimization: it rehearses future scenarios in virtual space before selecting the optimal strategy to execute in the physical world. As such, Ideal’s VLA, like Livis, fundamentally follows an embodied learning path, integrating perception and action within a single model. By contrast, Huawei’s world‑model approach resembles a safety‑first architectural paradigm: first simulating, then acting—essentially embodying a “generate–optimize–predict” workflow.

Whether it’s the smart cockpit, intelligent driving, or the in‑vehicle agent, both sides are deeply mired in the “sunk costs” and entrenched mindsets built up over years of technological investment, making any major shift extremely costly. This has turned the current competitive landscape into a one‑way bet, with vastly different stakes at play. For Huawei, its HarmonyOS‑enabled devices have already surpassed 1 billion units, and automobiles represent just one piece of its broader IoT ecosystem. Even if the in‑vehicle agent fails to gain traction, its massive ecosystem base is sufficient to dilute the risks. Moreover, its ADS system and HarmonyOS‑based cockpits have already been deployed across dozens of vehicle models, with a validation sample size far exceeding that of any single brand.
By contrast, the ideal bet is far more focused. It seeks to position AI as the cornerstone of its brand’s competitive moat. If Agent turns out to be a “false need,” failing to drive users’ willingness to pay or influence their purchasing decisions, then Li Auto’s massive investments in embodied intelligence could face a severe return‑on‑investment crisis. As the orchestrator, Li Xiang has repeatedly painted the vision of an “automotive robot that can charge and wash itself,” and the L9 Livis—its nascent embodiment—must deliver without fail. So does this mean Li Auto stands no chance? Not at all. Its true objective may not be to compete with HarmonyOS on the breadth of its application ecosystem, but rather to build a sufficiently deep moat in the realm of “physical intelligence.” The strategy is to get ahead of HarmonyOS—before it completes its transition from “ecosystem breadth” to “experience depth”—by helping users perceive and embrace the unique value of the “whole vehicle as an Agent,” thereby establishing robust cognitive barriers.

The key to whether Li Auto can successfully build an AI‑driven competitive moat likely lies in the time window—namely, swiftly enabling users to recognize and appreciate the unique value of Li Auto’s Agent. As with any new innovation, people tend to remember the first to propose it and put it into practice, and general‑purpose Agents are no exception. By leveraging this advantage, Li Auto’s viability as an automotive brand is assured. However, if Huawei’s HarmonyOS‑based Agent ecosystem achieves lock‑in before then, Li Auto’s differentiated user experience could face the risk of being matched and eroded.
Of course, even under the most optimistic scenario, Li Auto cannot upend the HarmonyOS ecosystem. What it can aim for is to offer users a deeply integrated, potentially paid “full‑vehicle Agent” solution in niche use cases—areas where HarmonyOS falls short and where users demand an exceptionally rich physical interaction experience. This race against time demands that Li Auto maintain an extremely rapid R&D‑iteration cadence to sustain a sufficiently pronounced experience gap. Once HarmonyOS closes that gap in terms of user experience, its scale advantages will enable it to roll out at lower costs, eroding Li Auto’s differentiating edge. Consequently, this is far from a simple one‑on‑one showdown; it is a high‑stakes race against time and technological depth, one that all new entrants betting on proprietary AI systems must face. The competition for in‑vehicle Agents—the long battle from concept to end‑user experience—has only just begun.

Reposted from Sina Auto

Email:
info@adtfm.com

Phone:
021-20532053

Address:
2nd Floor, No. 150 Jindian Road, Pudong New Area, Shanghai