今年早些时候，退休人员和年轻人排长队安装AI智能体OpenClaw的场景在中国疯传。尽管政府警告存在网络安全风险，消费者仍争相尝试这项被长期炒作的AI领域“下一个大事件”：不仅能聊天，还能自主行动的工具。
Earlier this year, scenes of retirees and young people waiting in long lines to install the AI agent OpenClaw went viral in China. Even as the government warned of cybersecurity risks, consumers rushed to try the long-hyped Next Big Thing in artificial intelligence: tools that don’t just chat, but act on their own.
中国似乎有望引领全球AI智能体的发展浪潮。中国的软件工程师走在前列，打造出了Manus等产品——Manus曾被视为全球领先者，吸引力之大甚至引起了Meta平台的注意，直至北京叫停了这笔收购。需求显然存在：根据Worldpay9月发布的调查，86%的中国消费者表示愿意用智能体购物，而美国这一比例仅为45%。
China seemed poised to lead the global push into AI agents. Its software engineers were ahead of the curve, producing products such as Manus, once regarded as a global leader and compelling enough to attract the attention of Meta Platforms Inc. before Beijing unwound the acquisition. The appetite is certainly there: 86% of Chinese consumers said they were willing to shop with one of them compared with just 45% in the US, according to a Worldpay survey released in September.
中国也具备技术能力。国内宠儿DeepSeek近期由其harness团队发布了一系列更新，该团队负责构建让AI模型更自主行动的工具。但有一个问题。这些努力仍高度集中于开发者而非普通消费者。而且理由充分。中国的消费级AI智能体被困住了。
China has the technical capability, too. Domestic darling DeepSeek has recently released a spate of updates from its harness team, which builds tools that allow AI models to act more autonomously. But there’s a catch. Those efforts remain heavily focused on developers rather than ordinary consumers. And for good reason. China’s consumer AI agents are trapped.
警钟随着Meta的Muse高调上线而敲响——Muse飙升至美国应用商店榜首，达到1000万美国下载量的速度甚至超过OpenAI的ChatGPT。自其发布以来的数周里，我听到大量困惑之声：为什么中国造不出同样能力的AI产品？现实是，技术确实存在，但仍被困在墙后。腾讯的微信和阿里巴巴的支付宝等超级应用在移动时代是令人生畏的护城河。
The wake-up call came with the buzzy launch of Meta’s Muse, which soared to the top of US app stores and reached 10 million US downloads faster than even OpenAI’s ChatGPT. In the weeks since its release, I’ve heard plenty of consternation about why China can’t build similarly capable AI products like this. The reality is that the technology exists, but it remains trapped behind walls. Super apps like Tencent Holdings Ltd.’s WeChat and Alibaba Group Holding Ltd.’s Alipay were formidable moats in the mobile era.
在智能体时代，它们开始更像围栏。曾经帮助建设中国互联网的架构，如今可能恰恰是阻碍它的东西。其数字世界仍被这些超级应用割据，而消费级智能体如果无法在它们之间自由穿行，就没有多大用处。
In the agentic era, they’re starting to look more like fences. The architecture that helped build China’s internet may now be precisely what’s holding it back. Its digital world remains carved up among these super apps, and a consumer agent isn’t much use if it can’t move freely between them.
这看起来并非主要是技术问题。字节跳动有限公司是中国最受欢迎的消费级AI平台“豆包”的制造商，已通过其“豆包手机”多次证明，智能体可以跨不同服务运行。问题在于权限。这些服务起初直接封杀了它，随后又拒绝授予访问权限。此类能力的发展已超越了管辖应用间交互的规则和标准。而掌握着数亿用户使用的服务的“守门人”不愿放弃这种优势，这也并不令人意外。
This doesn’t look primarily like a technical problem. ByteDance Ltd., maker of China’s most popular consumer AI platform Doubao, has repeatedly demonstrated through its Doubao phone that agents can operate across different services. The problem is permission. Those services first blocked it outright, then declined to grant access. Such capabilities have raced ahead of the rules and standards governing how apps interact. And it’s hardly surprising that the gatekeepers of services used by hundreds of millions of people are reluctant to surrender that advantage.
北京希望改变这一现状。在5月发布的一份政策文件中，官员呼吁制定智能体互联互通的关键标准和协议。但要打破这些壁垒，需要进行大规模的监管变革。6月也出现了裂痕，腾讯表示正与主要智能手机制造商合作，推行早期且受控的智能体间通信标准。没有哪里的墙比微信更高。我曾撰文指出，这个拥有14亿用户、实际上充当中国日常生活数字基础设施的平台，拥有竞争对手在竞争激烈的AI领域只能梦寐以求的护城河。但如果这道保护屏障变成了笼子，会怎样？
Beijing wants this to change. In a policy document released in May, officials called for the development of key standards and protocols for the interconnection of agents. But bringing down those walls will require a much larger regulatory shakeup. Cracks also appeared in June, when Tencent said it was partnering with major smartphone makers with early and controlled agent-to-agent communication standards. Nowhere are the walls higher than for WeChat. I’ve written before that the platform used by 1.4 billion people, effectively serving as digital infrastructure for daily life in China, has the kind of moat that competitors can only dream of in a hyper competitive AI sector. But what if that protective barrier has become a cage?
眼下这未必是坏事。自主系统或许受益于严格的管控，以免因意外行为让我们措手不及。大型科技公司——以及北京——都不愿失去控制权。围绕信任、责任和网络安全的问题依然笼罩着应用落地，更不用说在疫情以来青年失业率持续高企的经济环境下，对就业的长期风险了。超级应用在释放智能体时审慎行事是明智之举，但克制只能是暂时的特性，久了便会变成缺陷。
That might not necessarily be a bad thing for now. Autonomous systems could benefit from tight leashes to avoid them from surprising us with unintended behavior. Big tech companies — and Beijing — have no desire to lose control. Questions around trust, liability and cybersecurity still loom over adoption, not to mention the longer-term risks to jobs in an economy that has been grappling with persistently high youth unemployment since the pandemic. Super apps are wise to move slowly when unleashing agents, but restraint can be only be a feature for so long before it becomes a bug.
这种互操作性问题并非中国独有。我们已经看到，从大型银行到零售巨头亚马逊公司，都在其平台上屏蔽了部分智能体活动。但中国的超级应用生态系统让赌注大得多。只要在中国大陆待上一段时间，你就会迅速意识到，不经过微信或支付宝，很难完成许多事情。
This interoperability problem isn’t unique to China. We’ve already seen everyone from major banks to retail behemoth Amazon.com Inc. block some agentic activity on their platforms. But China’s super-app ecosystem makes the stakes far higher. Spend any time in the mainland and you’re quickly reminded how difficult it is to do much of anything without going through WeChat or Alipay.
OpenClaw 的兴衰暴露了其他障碍。智能体消耗的令牌和算力资源远多于聊天机器人，在免费 AI 服务泛滥的市场里，这让经济账算得更难看。在美国兴起的订阅制模式，历史上在中国软件消费者中一直难以获得认可。还有算力获取问题，这一直是悬在中国更广泛 AI 野心头上的制约。对于消费级智能体而言，经济账可能比技术更难算清。
OpenClaw’s rise and fall exposed other obstacles. Agents burn through significantly more tokens and computing resources than chatbots, making the math even uglier in a market already flooded with free AI services. Subscription models that have taken off in the US have historically struggled to gain traction among Chinese software consumers. Then there’s access to compute, a constraint that has hung over China’s broader AI ambitions. For consumer agents, the economics may prove harder to crack than the technology.
然而，腾讯在办公 AI 助手领域已占据领先地位，尽管它在消费级赛道上普遍被视为落后。据 QuestMobile 数据，其“微信小助手”工具月活用户已达约 3200 万，大致是最接近竞争对手的三倍。备受期待的微信智能体仍处于测试阶段，但腾讯可以复用推动“微信小助手”崛起的同一套打法和优势，来孵化面向大众市场的版本。
Yet Tencent has taken the lead in office AI assistants, even as it has largely been seen as falling behind in the consumer race. Its WorkBuddy tool has reached some 32 million monthly active users, roughly three times its closest rival, according to QuestMobile data. It’s long-hyped WeChat agent is still in beta, but Tencent could use the same playbook and advantages that propelled WorkBuddy’s rise to seed a mass-market version.
智能体仍需脚手架支持，而封闭生态系统可以提供这种支持。但超级应用推出自有系统或制定让他人接入的标准，时间不等人。否则，曾让它们城堡坚不可摧的高墙，恐将成为最大的束缚。
Agents still need scaffolding, and closed ecosystems can provide it. But the clock is ticking for the super apps to launch systems of their own or write the standards that let others in. Otherwise, the walls that made their castles so formidable risk becoming their biggest constraint.
尽管关于中国优势的炒作不断，到目前为止，它主要证明了成千上万的人愿意排队让 AI 智能体安装到设备上。硅谷则让数百万人无需如此。
For all the hype about China’s edge, so far it has mostly proved that thousands of people are willing to line up to have an AI agent installed on their devices. Silicon Valley made sure millions didn’t have to.