Dear Investors and Friends,

​​

“You won’t have to use different apps for different tasks. You’ll simply tell your device, in everyday language, what you want to do.” Bill Gates, essay on The Future of Agents

​

Digital platform power depends on monetising a scarce resource: user intent. Search engines monetise it through auctioning ad rankings, app stores through distribution fees, and super-apps through user time spent. We believe that in the future agentic operating systems (OS) on mobile phones have the potential to challenge digital platforms by becoming the gateway for users to express intent, and then controlling discovery, choice, and execution. This piece examines whether the potential introduction of powerful OS agents on mobiles represents a fundamental change in how digital platforms capture value.

​

Understanding the Shift: A Layman’s Guide to the Agentic OS

To understand the change we are describing, imagine your current smartphone as a digital shopping mall. To get anything done, you walk into a specific store (an app), find what you want, and check out. If you want to compare prices or book a full trip, you visit five different stores, carrying information between them in your head.

An agentic OS turns your phone from a mall into more of a personal assistant. Instead of you going to the apps, the OS does it for you.

​

What this could look like for the user if OS agents were fully unleashed: Instead of opening the Expedia app, searching for a flight, then opening the Calendar app to check your dates, and then opening WhatsApp to tell your spouse, you would simply hold the power button and say: "Book that London trip we discussed in WhatsApp for the second week of June." The phone could see the dates in your chats, find the flights in the background, and simply show you a confirm button. You would never actually open an app.

 

This shift in how you might use a mobile may seem incremental, but it has potentially game-changing implications for how digital platforms make money—as we will discuss.

 

1. The Global Landscape

What Google and Apple are doing: Google has repositioned its assistant layer — previously Google Assistant — around Gemini, a family of large language models that increasingly act as contextual AI agents. These agents are built to understand user intent across apps, handle multi-step tasks, and offer proactive suggestions. Apple’s AI initiative branded, Apple Intelligence, aims to evolve Siri into a system agent that understands user intent across apps and services. However, Apple has pushed back the full rollout for now.

​

Why don’t we have these powerful OS agents today? The technology is not really the problem, though of course everyone who’s used an AI chatbot knows it could be better. The issue is more that there are major regulatory and business hurdles to implementing a powerful OS agent.

 

The Success Trap: Apple and Google are constrained by their own success. These constraints are twofold: competition law and the innovator’s dilemma.

 

There may be a collision at some point between agentic OS design and competition law. An OS agent becomes a decision-maker when it sees all user content, decides which service should execute on user intent and then executes without the user manually choosing between competing options. Self-preferencing, which regulators watch out for, occurs when a gatekeeper platform advantages its own services over that of rivals. The regulatory standard is “Does this design structurally enable blocking competitors out of effective access to the user?” The OS agent cannot become the deciding economic actor on the device without being treated as a monopolistic allocator—so regulatory action is practically inevitable.

 

Additionally, agentic OS design threatens the app store tax with an innovator’s dilemma. Today, Apple and Google extract a distribution fee on third-party digital commerce via their stores. The economics are amazing: 15–30% take rate on in-app purchases, subscriptions, and digital goods with near-zero marginal cost. We are talking about tens of billions of dollars per year in earnings across the duopoly. With agentic OS, the agent disintermediates the app but also the purchase. Powerful OS agents would not just gut the high margin service revenue, but also likely severely damage the app developer ecosystem. The duopoly cannot simply move the tax to the OS agent by saying any transaction executed by the OS agent is taxed without collapsing the legal, economic, and ecosystem equilibrium that makes the app store tax viable in the first place. Do Apple and Google really want to cannibalise their highest margin business?

 

Beyond these two constraints that stem from the duopoly’s success, there is a further challenge at the payment rails level. Payment systems today do not recognise third-party delegated agents. Issuers, networks and payment processors lack a framework to identify the agent as the initiating party, assign liability, price risk, or resolve disputes. This is an important reason why there aren’t yet examples of third-party agents completing transactions on behalf of consumers.

 

Google, which is more aggressive on agentic OS than Apple, is currently shipping phones with Gemini-powered AI features pre-installed. But it’s more at an early hybrid stage where AI enhances specific workflows but does not replace the traditional app interface. Critically, Google has not yet enabled Gemini to orchestrate 3rd party apps, much less transact.

 

2. Imagining the Disruption

It’s important to think through what the scale of the disruption could look like if OS agents develop and become powerful (if the regulatory and business hurdles were somehow overcome). Perhaps the largest impact is the shift as the ecosystem moves from browsing (high value for ads) to delegating (zero value for ads).

 

Travel & Logistics: Instead of opening the Booking.com app to compare hotels, you tell your mobile: "Find me a 4-star hotel with a gym in London for under $300." The OS agent scans multiple platforms behind the scenes, presents a single best option, and handles the checkout. Booking becomes a back-end fulfilment engine. The implicit margin it earns from discovery goes away. Or, even worse for Booking, perhaps the agent just checks hotel websites and books directly.

​

E-commerce: You need to replace a broken coffee maker. Instead of searching Amazon and scrolling through sponsored products, the agent analyses your purchase history and current reviews across the web. It buys the product on the platform with the fastest delivery, potentially even from the brand’s website. Amazon’s sponsored listings become irrelevant.

​

Local Services: You tell your mobile that you're hungry for sushi. The agent doesn't open DoorDash or Uber Eats; it checks your past preferences and local availability, routes the order to the restaurant's own interface, and simply notifies you when the delivery is five minutes away.

 

Thinking more fundamentally, the potential disruption is not disintermediation of apps, but disintermediation of aggregation itself.

​

3. China: a less predictable agentic AI environment

China has a very different mobile and app ecosystem from the rest of the world, and it is possible that the debate around agentic AI on mobiles, OS agents and app businesses plays out very differently.

​

Bytedance’s attack: In China, Bytedance recently created a limited release mobile that tests the boundaries of what is possible with app orchestration. The late 2025 launch of the Doubao Phone (a collaboration between ByteDance and hardware manufacturer ZTE) was an aggressive early experiment in pre-installed mobile agents. The Bytedance Doubao Phone does not rely on apps to provide permission to work. Instead, it uses a graphical user interface (GUI) agent—an agent that literally looks at the screen, identifies buttons, and clicks them just as a human would, but faster. This is different to an OS agent that has privileged access to apps and execution authority. A GUI agent is really an adversarial automation operating on top of existing interfaces. Doubao is interesting because it avoids app permissions, but that also means it is possible for apps to resist.

Just a bit more about the Doubao phone: it runs a split-inference architecture, both local and on-cloud. Simple tasks happen on-device using a 7B-parameter model for privacy and speed. Complex tasks are offloaded to the cloud.

​

Unlike a standard mobile, the Doubao agent has system memory. It might remember that you prefer aisle seats, or that you usually order coffee at on the way to work. It can use this context to make decisions across different apps.

​

Application ecosystem defence: The pilot release of the Doubao device triggered a defensive reaction from China’s largest platform companies, particularly Tencent and Alibaba. These firms rightly recognised that a powerful agent such as Doubao posed a structural threat. By inserting itself between the user and existing applications, the agent risked diverting attention, intent, and behavioural data away from the super-apps that rely on user interaction for monetisation and control.

​

The response was a series of defensive measures effective against the un-permissioned Duobao GUI agent. Some major apps began to defend their interfaces against GUI agent-driven interaction. They changed how their apps work so that a GUI agent could no longer move through them as easily as a human. They stopped clearly labelling buttons in ways that GUI agents rely on, and they added more checks on the backend to confirm that a real person, not an agent, was behind each action. If the agent clicks something slightly out of order, or differently to the way a human would, the action might simply fail. These steps did not fully block the Doubao GUI agent, but they deteriorated the user experience.

​

On 15 Jan 2026, Alibaba released what is in effect its answer to the challenge posed by Doubao. Ali beta released an update to its Qwen AI chatbot that features agentic orchestration within the Ali ecosystem. Unlike Doubao, Qwen does not operate as an un-permissioned GUI agent nor rely on OS-level control. Instead, Qwen acts across Ali’s own application ecosystem, including Taobao (commerce), Feizhu (travel), Ele.me and Shangou (local delivery), Gaode (mobility), and Alipay (payments). Users can pay without ever leaving the Qwen interface. Whether Ali’s countermove is decisive, or OS agents eventually triumph, is unclear today.

 

4. China’s OEM-Led Model vs. The DM Duopoly

The situation in China is different to the ex-China Google/Apple duopoly. China’s unique circumstances create a much wider range of outcomes for how OS agents develop in China.

​

Most Chinese mobile OEMs except Huawei (Huawei stopped using Google’s version of Android because of US sanctions) use Android-based systems that are heavily adapted. Google Mobile Services are removed from these devices. This means that the Gemini OS agent is not on Chinese Androids. Chinese OEMs (Huawei, Xiaomi, Oppo, Vivo) may each install their own OS agents, often based on 3rd party Chinese models. The Chinese OEMs also aren't as beholden to an app store revenue model (app stores in China have low take rates and are fragmented), and they may be more willing to do things differently to deliver a better user experience and hopefully win hardware market share.

​

While Western regulators are focused on fairness and interoperability, Chinese regulators may be more focused on national efficiency and system-level coordination. It’s possible that the Chinese government mandates that super-apps open their interfaces to OS agents in the interest of boosting national AI productivity. One result would be that people living in China would have much more useful mobile phones.

Another possible result of OS agents developing in China might be the disruptive de-moating of Tencent and Alibaba. But this is far from certain. It’s also possible that Ali’s Qwen AI Assistant is compelling for consumers and becomes an important entry point. In effect, Chinese super apps might internalise the agent layer rather than being disintermediated by it.

​

5. Variis Strategy Analysis: Digital Platform Disruption

From an investment perspective, the relevant question is not whether AI agents are technologically possible, but whether they are rolled out, and then which parts of existing businesses they might disrupt. Powerful OS agents could disrupt intent discovery, which simply means finding out what the user wants. They may be less disruptive when the crux is real-time physical coordination (ride hailing), high service complexity (“my flight got cancelled”), or trust-heavy human work (recruiting).

 

Using this framework, we bucketed our digital platform investments—MercadoLibre, Trip.com, Didi, Grab, Allegro, Coupang, Sea, Full Truck Alliance, Kanzhun, and MakeMyTrip—according to the functions they perform and analysed how exposed that function is to OS agent disintermediation.

​​

​​

​​

​​

​​

​​

​

​

​

​

​

 

 

 

For discovery platforms like marketplaces, assuming powerful OS agents were rolled out, it is possible that over time monetisation could shift from ads to OS agent access pricing, almost like a credit card processor fee. Fulfilment seems less vulnerable from our perspective.

​

6. Strategy Summary

OS agents raise the question of how value will shift between businesses. If intent expression, discovery, and task execution move to OS agents, then platforms that monetise attention may be disrupted. By contrast, platforms involved in real-time physical coordination, high service complexity, or trust-heavy human work may be safer.

Outside China, we expect any shift to powerful OS agents to be constrained. For Apple and Google, considering potential regulatory scrutiny as well as the app-store revenue model, OS agent orchestration and transaction execution is unlikely to proceed quickly. Also, platform businesses will resist.

​

China has a more uncertain range of outcomes. Agentic disruption could move faster. Our strategy positioning reflects this asymmetry. We have limited exposure to ad-driven discovery platforms in China. Instead, our China platform investments are concentrated in businesses focused on real-time physical coordination, high service complexity, and trust-heavy human work. Our Chinese OTA business Trip is potentially riskier in our assessment, and to reflect that we have reduced the position size. We continue to believe that Trip’s contracted hotel inventory and high customer service strategy create an effective defence against disintermediation risks.

 

Our analysis of China’s digital ecosystems is grounded in a bottom-up understanding of these businesses. This includes paying close attention to platform experiments and the incentives driving behaviour across mobile OEMs, app platform businesses and other stakeholders. Years of investing across Emerging Markets, combined with Chinese-language capability, extensive time on the ground, and deep local networks, give us a better chance of assessing these dynamics accurately. We don’t think these things are easy to understand from a distance.

 

Thank you for your continued interest and support!

Leila, Eko, Rufus and Jamie

​​

Disclaimer

FOR PROFESSIONAL INVESTORS AND ADVISORS ONLY
The contents of this document are communicated by, and the property of, Variis Partners LLP. The information and opinions contained in this document are subject to updating and verification and may be subject to amendment. No representation, warranty, or undertaking, express or limited, is given as to the accuracy or completeness of the information or opinions contained in this document by Variis Partners LLP or its directors. No liability is accepted by such persons for the accuracy or completeness of any information or opinions. As such, no reliance may be placed for any purpose on the information and opinions contained in this document. The information contained in this document is strictly confidential and is not intended to be advice or an offer or solicitation to invest. The value of investments and any income generated may go down as well as up and is not guaranteed.