Is the Chinese version of Muse worth betting on for BATs?

The moves of BAT themselves have given the answer: copying Meta's playbook blindly will inevitably lead to poor adaptation to local conditions, but in the field of personal AI agents, strategic positioning is a must.

After Muse, the personal AI agent launched by Meta, became a hit, Chinese tech companies have also sprung into action. ByteDance ramped up R&D efforts before the National Day holiday, while the Doubao team is advancing a personal assistant project codenamed "Spell" in the form of an independent app. Tencent launched LightVela, a cloud-based personal assistant, at the end of September, and is currently conducting internal testing of Handy Bot, a personal agent product running as an independent app.

Muse gained 5 million downloads just 3 weeks after its launch, driving giants including OpenAI, Google and Anthropic to follow up across the board. Against this backdrop, it is not surprising that domestic tech giants have taken action promptly. "Silicon Valley takes the lead, and China follows" has already become the norm.

Here comes the question: From OpenClaw to Vibe Coding to AI office scenarios, the focus of AI competition has kept changing since the beginning of this year. Is it still necessary for BAT to dance to Meta's rhythm?

The moves of BAT themselves have given the answer: copying Meta's playbook blindly will inevitably lead to poor adaptation to local conditions, but in the field of personal AI agents, strategic positioning is a must.

Many people assume that BAT is aiming to build a "Chinese version of Muse", but this ignores the differences in implementation paths. Both are personal AI agents, but Muse adopts an independent cloud virtual machine architecture, focusing on autonomously completing general tasks; while BAT is taking the route of integrating existing ecosystems, tending to embed agents into their existing service systems, with significantly more restrained permission control.

The reason is simple: the "internet landscape" is different. The reason why Muse can operate smoothly in the United States largely benefits from the relatively open web ecosystem. Meta can access Google Workspace thanks to public APIs and an open authorization protocol system. However, in China, top apps are fragmented, making it almost impossible for an agent to read WeChat chat records, call Douyin's recommended content and place orders on Taobao at the same time.

At present, the phenomenon of "agent islands" is very obvious in China: Alibaba will encapsulate its app ecosystem into skills that can be called by its own agents, ByteDance's ArkClaw will deeply integrate Feishu and Douyin plugins, and Tencent will also focus on building its own SkillHub (AI skill community). The possible scenario in the next stage is that those super apps will each host their own personal AI agents, and there is little chance that a mirror version of general-purpose personal agents like Muse will emerge in China.

Although it is impossible to replicate Muse, BAT has no choice but to "call" in this game. This is not so much following the trend as avoiding missing the window of paradigm shift and making strategic moves in the competition for the future.

Why has Muse become so popular? To a large extent, it is because it is expected to reconstruct the super traffic entry.

Over the past 20 years, search engines, social platforms and super apps have successively become iconic super traffic entries. As the era shifts to the AI era, the hope of carrying the next-generation super traffic entry falls on AI agents that can understand user needs, call cross-platform services and automatically complete complex tasks.

At present, personal AI agents represented by Muse run on cloud persistent virtual machines, which are very suitable for undertaking long tasks spanning multiple periods and multiple steps. They go significantly further than previous agents that can execute autonomously across platforms but are limited to single instructions and reset sessions once tasks are completed, making them more like "the next super entry".

As personal AI agents are linked to the imagination of "the right to allocate future traffic", failing to seize the entry may degenerate enterprises into "back-end service providers". It is therefore reasonable for BAT to lay out personal AI agents.

However, it should also be noted that as the daily new users of Muse have slowed down significantly, the path ahead for personal AI agents is not all smooth. Issues including privacy protection and profit model may become key constraints for their popularization.

The value and risks of personal AI agents both stem from one point: they log in for you, make payments for you, and make decisions for you. Once users grant permissions to the agent through "one-stop authorization", they will face problems such as vague attribution of responsibility.

Meta's solution is to store credentials separately, monitor through a "sentinel" mechanism, and require manual confirmation for sensitive operations. However, surveys show that only 8% of American users are willing to give their passwords to Meta for agent operations. Two weeks after Muse's launch, a "zero-day vulnerability" was exposed, which means a simple attack can completely hijack the agent and access all authorized platform permissions, including shopping accounts and payment information. This also shows that it still has a long way to go before achieving full reliability.

Muse currently offers a free basic version and a paid version. Meta executives said that "it may eventually take a cut from transactions", but this calculation is more complicated than billing by token usage.

Different from conversational AI which mainly consumes GPU inference resources, the multi-step tasks of agents require browsers, API calls, virtual machines and task scheduling. CPU is becoming the computing power base for large-scale operation, which will inevitably continue to increase the operating cost per task. Whether the profit model can work is still in doubt at the moment.

Considering factors such as possible narrowed permissions and the end of users' novelty-seeking period, it can only be said that the popularity of Muse does prove the existence of demand, but the trust gap is not easy to cross, and the profit model is not easy to verify either. For this reason, Meta's internal attitude towards Muse's test is "mixed", and domestic manufacturers generally choose to test the water through "developer experience". All these indicate that personal AI agents are still far from being truly popularized among the general public.

In the long run, the trend that the next "internet entry" will be agents is irreversible. The party that can strike a good balance between security, privacy, openness and user experience is more likely to win the competition for the entry. In this regard, what Muse has presented is not so much an answer as an examination question left for tech enterprises to answer.

This article is from the WeChat official account "Economic Observer", written by She Zongming, and published with authorization from 36Kr.