For an AI agent to connect a car's many functions and keep improving after the sale, the traditional model of one-time development and delivery is no longer sufficient. Desay SV is building toward a "super agent" that ties driving, vehicle control, cockpit, and cloud into a single user experience — a shift that asks automakers and Tier 1 suppliers to decide not just what new features to add, but what they design together and how far their responsibility extends.

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기







Desay SV's Zhou Jian presented "AI Agent-Led New Paradigm of Automotive Value: From Parameter Competition to Resonance of Experience," arguing that intelligent-vehicle competition is shifting from specifications to user experience.




Cars already do a great deal. Large displays, high-performance chips, cameras and sensors, voice assistants and app suites are no longer reserved for a handful of premium models. Screens have grown, computing power has climbed, and the list of things a car can do keeps getting longer. But when similar hardware and software show up across brands within a single product cycle, specifications stop explaining much on their own. A feature that felt new the first time rarely feels new the fifth.
Specs keep rising, but the sense of something genuinely new keeps shrinking. The problem isn't a shortage of features. Seats, climate, lighting, music, navigation, and external services can each work fine on their own and still add up to nothing coherent if they don't move together around what the person in the car actually needs at that moment.
Zhou Jian of Desay SV described the next stage of this competition as "resonance of experience" — a car that understands a passenger's state and situation, and pulls scattered functions together at the moment they're needed. In his account, the first half of the intelligent-vehicle race was fought over parameters and feature counts; the second half will be fought over user experience.
If a passenger is tired, the car should be able to adjust the seat, climate, lighting, and music at the same time. If arrival is running late, the car shouldn't simply recalculate the route — it should also adjust the user's schedule and the destination-side services tied to that trip. The point isn't adding more functions. It's deciding, on the car's own initiative, who those functions serve and when they should act together. That's the lens Desay SV is applying to AI agents.



Not More Functions, but Better Coordination

Most in-car voice assistants today work by recognizing what a person says and running a matching command. They can turn on the air conditioning, search for a destination, or play music, but they struggle to read the relationship between several requests, sequence the steps, or follow through on the result. Reading a situation the driver hasn't put into words — and acting on it before being asked — is harder still. Zhou was clear that an agent, in his framing, is not a chatbot that answers one question at a time.
What Desay SV describes is an agent that reads intent, plans the steps needed to satisfy it, and calls on functions and services inside and outside the car to carry it out. Ask whether there's a coffee shop nearby, and instead of stopping at a list of results, the agent checks the destination, arrival time, and preferences, picks a suitable place, and carries the request through ordering, payment, and pickup timing.
Making that happen draws on far more than a single voice model. Cameras, microphones, and vehicle controllers capture information about the occupants, surroundings, and vehicle state. On-board voice recognition and data services handle the immediate response and basic control. Complex queries, multi-intent reasoning, knowledge lookup, and long-term memory sit with the cloud and large AI models. A vehicle-control agent, a mobility agent, a knowledge agent, a lifestyle-services agent, and a decision agent each handle their own piece while working toward one outcome.
Memory is what turns a general-purpose feature into something personal. Storing a single conversation isn't enough for a car to adapt to someone over time. Short-, medium-, and long-term memory need to connect so that preferences, habits, and recurring travel patterns accumulate — otherwise the car keeps asking the same questions instead of making better calls. If someone drives the same route every weekday morning and prefers the same temperature and playlist, the car should be able to check the current situation and offer those settings before being asked, rather than waiting for a command each time.
None of that means every decision can be handed to the cloud. A car still has to handle driving and safety-critical functions in areas with poor connectivity, including underground roads and parking facilities. A slow response from the cloud can't turn into a slow response from the vehicle's controls. Desay SV's answer is a division of labor: on-board AI handles immediate response and core control, while the cloud takes on complex tasks, external services, long-term memory, and continuous learning. The large AI model sits between the two, tying them together, broadening the agent's capabilities, and accelerating feature development.
Under this arrangement, the car stops being a tool that waits for one instruction at a time. It becomes a system that reads a person's intent and situation together, chooses what to do, and checks the outcome — a shift from a passive tool waiting to be told, to something closer to a companion that understands first and acts on that understanding.




Desay SV's proposed AI agent architecture divides responsibilities between the vehicle and the cloud. On-board AI handles immediate responses and core control, while the cloud supports complex tasks and a broader service ecosystem. Large AI models and short-, medium-, and long-term memory expand the system's capabilities and support continuous development.



Beyond the Cockpit, Across the Whole Car

For an AI agent to coordinate the whole vehicle, it can't stay confined to the cockpit. Cockpit teams and intelligent-driving teams are still, in most companies, organized and engineered separately — one handling displays, infotainment, voice, and apps, the other handling sensing, perception, decision-making, and vehicle control. But for the person inside, mobility and the in-cabin experience are not separate.
If the intelligent-driving system reads traffic and road conditions and sees that arrival will be delayed, a cockpit agent can use that to adjust the calendar or suggest destination-side services. If fatigue is detected, the same system can adjust seat, climate, lighting, and music while also prompting the driver to rest and triggering the relevant safety response. Deciding when and where to charge needs the battery level, charger congestion, and the driver's schedule considered together.
The interface changes with it. Instead of the same menu and the same set of apps for every driver, the screen becomes something that surfaces what's relevant to the current situation and intent first. The driver no longer has to remember where a function lives or how to reach it; the agent finds it and presents it in whatever form fits.
What Desay SV describes as its end goal is a "super agent" — a system with unified perception, unified decision-making, unified execution, and continuous learning. This isn't one giant model doing everything itself. It's a layer that coordinates agents across different domains, along with vehicle functions and external services, and delivers a single result to the person in the car.
Building that requires more than the agent software itself. It needs a computing platform to run AI in the car, an AI OS to connect functions and services, and an "AI box" that can bring AI capability to existing vehicles and different electronic architectures. It also needs a development process that carries AI all the way from design and validation through production and post-sale updates. That's why Desay SV treats AI agents as a shift in the vehicle platform itself, not as another cockpit feature.
This is also where Desay SV's existing business lines reconnect. Supplying cockpit displays, domain controllers, intelligent driving, and vehicle networking as separate products doesn't add up to one experience on its own. Information generated in one domain has to feed a shared decision, and that decision has to be carried out across vehicle control and external services alike. The AI agent becomes the point where previously separate technology areas meet, organized around the experience rather than the component.




Desay SV has widened its collaboration with OEMs across four stages — features, platform, user experience, and AI capability — starting with a co-developed cockpit platform in 2019, customer-specific cockpit features in 2023, multi-customer interaction optimization in 2024, and a joint AI lab with Alibaba Cloud established in 2026.



From Supplier to Co-Designer

What AI agents change isn't only how the car behaves. It also touches how people pay for one. Zhou expects the revenue model to move away from a single purchase and toward software updates that add features and benefits gradually after the sale.
That shift reaches into how automakers and suppliers work together. In the traditional structure, the OEM sets requirements and the supplier builds hardware and software to match, delivering it once. A supplier's responsibility and business opportunity were built around the start of mass production. An AI agent has to keep learning from data and adding features and services after the car is sold. As how people use the services and respond to them changes, the agent's judgment and interface need to change with it. Every new external service or software update means the whole experience has to be validated again. A one-time delivery model that separates development, production, and post-sale operation into different phases can't keep up with that. Zhou went further, arguing that the AI agent era calls for long-term co-development and building capability jointly, not simply working more closely.
Desay SV has widened the scope of that collaboration across four stages: features, platform, user experience, and AI capability. It started by building individual features to a customer's specification, moved to building the shared platform that cockpit and agent run on, and is now defining, together with automakers, the actual moments a driver will experience before the underlying technology is even built.
That progression happened in steps. In 2019, Desay SV co-developed a cockpit platform with a customer. In 2023, it built cockpit features for a specific customer. In 2024, it optimized interaction experiences across several customers. In 2026, it set up a joint AI lab with Alibaba Cloud to develop large AI models, cloud capability, and in-vehicle AI together.
What matters in that timeline isn't the number of joint projects. It's that Desay SV's role has moved from implementing individual features, to co-developing the platform, to jointly defining the user experience and jointly building AI capability. As AI agents connect more of the car and keep changing after the sale, a supplier can't stay in the business of building to spec alone. Desay SV is positioning itself as a co-designer that works with automakers from the earliest concept and keeps improving the experience through data and updates long after production starts.
None of this erases the line between OEM and Tier 1 overnight. If anything, the more vehicle functions and data an agent connects, the more precisely that line needs to be drawn. That raises several questions: Who manages the user's data and long-term memory? Who controls the interface? Who designs and validates the authority connecting vehicle control with lifestyle services? If an agent's judgment reaches into a safety-related vehicle action, responsibility for errors and accidents has to be spelled out just as precisely.
Revenue-sharing is its own unresolved question. How value generated by post-sale services gets divided among the OEM, the Tier 1, the cloud and model providers, and external service operators will shape what the partnership actually looks like. Agreeing in principle that the ecosystem should be open is easy. Actually dividing up data, customer access, and decision-making authority is harder than any technical integration.
Competing on AI agents won't be settled by a better voice assistant or a larger AI model. It requires managing driving, vehicle control, cockpit, and cloud as one experience across a car's entire life. OEMs need a way to work with outside capability while keeping control of the vehicle and the brand; Tier 1 suppliers need software and data capability that extends well past the point of delivery.
Zhou's "resonance of experience" comes back, in the end, to a question about how the industry divides its work: who pulls scattered functions into one experience, and who keeps improving it after the car is sold. The answer will determine whether Tier 1 suppliers remain implementers — or become co-designers in the AI agent era.



Desay SV envisions a vehicle-wide intelligent architecture built on AI Agents as a unified foundation, enabling unified perception, decision-making and execution as well as continuous evolution. It integrates the cockpit and intelligent driving to deliver cross-domain experiences and dynamically generated personalized interfaces.

AEM(오토모티브일렉트로닉스매거진)