Meta Platforms’ breakout Muse personal AI agent is shining a spotlight on a new battleground in artificial intelligence – one in which China’s biggest internet companies are already jostling for position.
Launched on September 8, Muse surpassed 2.5 million downloads in its first two weeks and topped free-app rankings on Apple and Google’s US app stores. Its popularity has drawn fresh attention to the emerging market for what are known as agent harnesses.
Unlike chatbots, which merely respond to prompts, agent harnesses provide underlying AI models with persistent memory, tool access and the autonomy to execute more complex tasks.
Meta says Muse operates across the web and connected services like email and calendars, and can help users fill out forms, make purchases and book travel while remembering their preferences.
Industry analysts said agent harnesses could shift more of the AI sector’s value away from foundational models and towards the software layer.
Analysts at financial research and brokerage firm Bernstein said in a recent report that the main prize was “context lock-in”: the more an agent learned about a user’s habits, history and workflows, the more useful – and harder to replace – it could become.
That dynamic could favour Chinese internet giants Tencent Holdings and Alibaba Group Holding – the owner of the South China Morning Post – whose sprawling ecosystems encompassed messaging, payments, e-commerce, workplace software and cloud computing, the report said.
Muse’s success had revived investor interest in the value of such integrated ecosystems, said Winnie Wu, head of Asia-Pacific equity strategy and co-head of China equity research at Bank of America.
“With a chatbot, it doesn’t have a differentiation point,” Wu said. “It’s about whether the agent can get things done for you.”
Local front-runners
As the paradigm shifts, Chinese tech heavyweights are deploying their own agent strategies to capture market share.
Tencent stood out on the consumer side, Bernstein said, with its WorkBuddy workplace agent having taken an early engagement lead among third-party productivity harnesses in China, ahead of rivals including ByteDance’s Trae and Alibaba’s QoderWork.
The analysts said Tencent could take “a small leap of faith” towards building something similar to Muse by making its WeChat social media platform the front-end interface for the operation of its Xiaowei agent behind the scenes.
Xiaowei, which entered beta testing in June, is designed to execute tasks through WeChat mini programs rather than simply answer questions.
By leveraging native access to WeChat search, mini-programs and WeChat Pay, Xiaowei could turn Tencent’s ecosystem – which processes trillions of yuan in annual transactions – into a powerful data flywheel, the Bernstein report said.
Deepening agent usage would lock users further into Tencent’s services while generating high-quality, real-world data to refine future AI models, it said.
The Bernstein analysts said Alibaba was pursuing a more enterprise-focused strategy, integrating agents across QwenWork, DingTalk, Taobao and Alibaba Cloud.
The company could use existing enterprise data and workflows to embed agents more deeply into corporate operations, they said, with greater adoption to drive demand for its Qwen model and cloud computing services.
ByteDance could emerge as a contender through Feishu, its workplace collaboration platform. Bernstein said the company’s hyperscaler capacity could allow it to expand further into enterprise agents, but it had yet to demonstrate the same ecosystem position as Tencent in social communications or Alibaba in enterprise software.
Competition is also emerging outside China’s internet giants, particularly in coding. Bernstein said DeepSeek Harness and Z.ai’s ZCode ranked eighth and eleventh respectively among the coding tools it tracked on OpenRouter by 30-day token usage, although global players such as Anthropic’s Claude Code remained well ahead.
Manus, the Chinese-founded agent start-up that briefly became part of Meta before resuming independent operations, is also moving closer to the personal-agent model.
Last week, it unveiled Manus 2.0, including a new agent harness called Cascade, and Cue, a dedicated personal-agent app.
Analysts said the emerging competition suggested the AI race could increasingly move beyond model performance to models that know their users best.
Hurdles in China
However, Bank of America’s Wu cautioned that the Muse model might be harder to replicate in the Chinese market.
She said China’s mobile internet was dominated by “guarded” ecosystems that could make it difficult for an agent to operate seamlessly across services. Tasks involving WeChat or Alipay, for example, might require access that an agent controlled by a rival platform could not easily obtain.
Consumer behaviour posed another challenge. Chinese users had historically shown less willingness than their counterparts in the United States to pay recurring subscriptions for software, Wu said, raising questions about how personal agents would be monetised.
There was also the cost of running them. Agents such as Muse could require substantially more computing power – a constraint faced by China – than conventional chatbots, Wu added. -- SOUTH CHINA MORNING POST
