Aditya DidwaniyaMeta’s Muse app triggered a new AI contest. Analysts say AI agent harnesses, not chatbots, could decide who wins the next phase of artificial intelligence.
A New Kind of AI Contest Is Taking Shape
Meta’s personal AI assistant, Muse, did something few apps manage these days. It crossed 2.5 million downloads within two weeks of its September 8 launch. It also topped the free app charts on both Apple’s and Google’s US stores.
That kind of traction has put a spotlight on a part of AI that most casual users have barely noticed: agent harnesses. These are different from the chatbots people are used to. A chatbot answers your question and waits for the next one. An agent harness gives an AI model memory, access to outside tools, and the ability to act on its own across several steps.
Meta describes Muse as something that can move across the web and connected apps like email and calendars. It can fill out forms, make purchases, and book travel. It also remembers user preferences over time, so it gets more useful the longer it’s used.
Why Analysts Are Watching This Closely
Brokerage firm Bernstein recently argued that the real prize in this race isn’t the smartest model. It’s what analysts call context lock-in. The idea is simple. The more an agent learns about someone’s habits and daily workflows, the harder that agent becomes to replace.
That logic could work in favor of China’s largest internet platforms. Companies like Tencent and Alibaba already run sprawling ecosystems that combine messaging, payments, shopping, workplace tools, and cloud services. Winnie Wu, who leads Asia-Pacific equity strategy at Bank of America, said chatbots alone don’t offer much to differentiate. She put it plainly: “””It’s about whether the agent can get things done for you.”””
China’s Internet Giants Are Already Moving
Tencent appears to be ahead on the consumer side. Its workplace agent, WorkBuddy, reportedly leads early engagement among productivity tools in China, ahead of ByteDance’s Trae and Alibaba’s QoderWork. Bernstein suggested Tencent could go further by turning WeChat into the front door for its Xiaowei agent, which has been in beta testing since June.
Xiaowei is built to complete tasks through WeChat’s mini programs, not just chat. Since WeChat already handles enormous volumes of daily payments and services, deeper agent use could feed Tencent a steady stream of real-world data while keeping users inside its ecosystem.
Alibaba appears to be taking a different route, focused more on business customers than individual consumers. It’s weaving agents into QwenWork, DingTalk, Taobao, and Alibaba Cloud. Analysts expect this enterprise push to drive more demand for Alibaba’s Qwen models and its cloud computing business.
ByteDance has its own card to play through Feishu, its workplace collaboration platform, backed by serious computing capacity. Still, it hasn’t yet built the kind of ecosystem reach that Tencent has in social messaging or Alibaba has in enterprise software.
Coding Agents and a Rising Challenger
Competition isn’t limited to China’s biggest names. In coding specifically, Bernstein ranked DeepSeek’s Harness and Z.ai’s ZCode eighth and eleventh among tools tracked by usage on OpenRouter. Global leaders like Anthropic’s Claude Code still hold a clear lead in that space.
Manus, the Chinese-founded agent start-up that briefly joined Meta before going independent again, is also pushing into personal agents. It recently launched Manus 2.0, featuring a new harness called Cascade and a standalone personal-agent app named Cue.
The Obstacles China Still Faces
Despite the opportunity, Wu warned that replicating Muse’s success in China won’t be straightforward. China’s mobile internet is split into separate, closely guarded ecosystems. An agent built by one company might struggle to operate smoothly inside a rival’s app, such as WeChat or Alipay.
User habits add another layer of difficulty. Chinese consumers have generally been less willing than American users to pay ongoing subscriptions for software, which complicates how personal agents could eventually make money.
Then there’s the computing cost. Agents like Muse typically need far more processing power than a standard chatbot, and that’s an added strain in a market already facing chip supply constraints.
Even with these hurdles, the shift in direction is clear. The next phase of the AI race may not be won by whoever builds the smartest model. It may go to whoever builds the agent that knows its users best.
Mayur Mohta
Akshay Didwaniya
