Final 12 months, China noticed a growth in basis fashions, the do-everything massive language fashions that underpin the AI revolution. This 12 months, the main target has shifted to AI brokers—techniques which can be much less about responding to customers’ queries and extra about autonomously carrying out issues for them. 

There at the moment are a bunch of Chinese language startups constructing these general-purpose digital instruments, which might reply emails, browse the web to plan holidays, and even design an interactive web site. Many of those have emerged in simply the final two months, following within the footsteps of Manus—a normal AI agent that sparked weeks of social media frenzy for invite codes after its limited-release launch in early March. 

These rising AI brokers aren’t massive language fashions themselves. As a substitute, they’re constructed on prime of them, utilizing a workflow-based construction designed to get issues accomplished. Quite a lot of these techniques additionally introduce a unique means of interacting with AI. Quite than simply chatting backwards and forwards with customers, they’re optimized for managing and executing multistep duties—reserving flights, managing schedules, conducting analysis—by utilizing exterior instruments and remembering directions. 

China might take the lead on constructing these sorts of brokers. The nation’s tightly built-in app ecosystems, speedy product cycles, and digitally fluent consumer base might present a good surroundings for embedding AI into every day life. 

For now, its main AI agent startups are focusing their consideration on the worldwide market, as a result of the most effective Western fashions don’t function inside China’s firewalls. However that would change quickly: Tech giants like ByteDance and Tencent are getting ready their very own AI brokers that would bake automation instantly into their native super-apps, pulling knowledge from their huge ecosystem of packages that dominate many features of every day life within the nation. 

Because the race to outline what a helpful AI agent seems like unfolds, a mixture of bold startups and entrenched tech giants at the moment are testing how these instruments would possibly really work in apply—and for whom.

Set the usual

It’s been a whirlwind few months for Manus, which was developed by the Wuhan-based startup Butterfly Impact. The corporate raised $75 million in a funding spherical led by the US enterprise capital agency Benchmark, took the product on an bold international roadshow, and employed dozens of latest workers. 

Even earlier than registration opened to the general public in Might, Manus had change into a reference level for what a broad, shopper‑oriented AI agent ought to accomplish. Quite than dealing with slim chores for companies, this “normal” agent is designed to have the ability to assist with on a regular basis duties like journey planning, inventory comparability, or your child’s college undertaking. 

Not like earlier AI brokers, Manus makes use of a browser-based sandbox that lets customers supervise the agent like an intern, watching in actual time because it scrolls by internet pages, reads articles, or codes actions. It additionally proactively asks clarifying questions, helps long-term reminiscence that may function context for future duties.

“Manus represents a promising product expertise for AI brokers,” says Ang Li, cofounder and CEO of Simular, a startup based mostly in Palo Alto, California, that’s constructing pc use brokers, AI brokers that management a digital pc. “I consider Chinese language startups have an enormous benefit in relation to designing shopper merchandise, due to cutthroat home competitors that results in quick execution and higher consideration to product particulars.”

Within the case of Manus, the competitors is shifting quick. Two of essentially the most buzzy observe‑ups, Genspark and Flowith, for instance, are already boasting benchmark scores that match or edge previous Manus’s. 

Genspark, led by former Baidu executives Eric Jing and Kay Zhu, hyperlinks many small “tremendous brokers” by what it calls multi‑element prompting. The agent can change amongst a number of massive language fashions, accepts each pictures and textual content, and carries out duties from making slide decks to inserting telephone calls. Whereas Manus depends closely on Browser Use, a well-liked open-source product that lets brokers function an online browser in a digital window like a human, Genspark instantly integrates with a wide selection of instruments and APIs. Launched in April, the corporate says that it already has over 5 million customers and over $36 million in yearly income.

Flowith, the work of a younger group that first grabbed public consideration in April 2025 at a developer occasion hosted by the favored social media app Xiaohongshu, takes a unique tack. Marketed as an “infinite agent,” it opens on a clean canvas the place every query turns into a node on a branching map. Customers can backtrack, take new branches, and retailer ends in private or sharable “information gardens”—a design that feels extra like undertaking administration software program (suppose Notion) than a typical chat interface. Each inquiry or activity builds its personal mind-map-like graph, encouraging a extra nonlinear and artistic interplay with AI. Flowith’s core agent, NEO, runs within the cloud and might carry out scheduled duties like sending emails and compiling recordsdata. The founders need the app to be a “information marketbase”, and goals to faucet into the social side of AI with the aspiration of changing into “the OnlyFans of AI information creators”.

What in addition they share with Manus is the worldwide ambition. Each Genspark and Flowith have said that their main focus is the worldwide market.

A world tackle

Startups like Manus, Genspark, and Flowith—although based by Chinese language entrepreneurs—might mix seamlessly into the worldwide tech scene and compete successfully overseas. Founders, traders, and analysts that MIT Know-how Assessment has spoken to consider Chinese language firms are shifting quick, executing effectively, and shortly developing with new merchandise. 

Cash reinforces the pull to launch abroad. Prospects there pay extra, and there are loads to go round. “You may worth in USD, and with the change price that’s a sevenfold multiplier,” Manus cofounder Xiao Hong quipped on a podcast. “Even when we’re solely working at 10% energy due to cultural variations abroad, we’ll nonetheless make greater than in China.”

However creating the identical performance in China is a problem. Main US AI firms together with OpenAI and Anthropic have opted out of mainland China due to geopolitical dangers and challenges with regulatory compliance. Their absence initially created a black market as customers resorted to VPNs and third-party mirrors to entry instruments like ChatGPT and Claude. That vacuum has since been stuffed by a brand new wave of Chinese language chatbots—DeepSeek, Doubao, Kimi—however the urge for food for overseas fashions hasn’t gone away. 

Manus, for instance, makes use of Anthropic’s Claude Sonnet—broadly thought-about the highest mannequin for agentic duties. Manus cofounder Zhang Tao has repeatedly praised Claude’s capacity to juggle instruments, bear in mind contexts, and maintain multi‑spherical conversations—all essential for turning chatty software program into an efficient govt assistant.

However the firm’s use of Sonnet has made its agent functionally unusable inside China and not using a VPN. When you open Manus from a mainland IP tackle, you’ll see a discover explaining that the group is “engaged on integrating Qwen’s mannequin,” a particular native model that’s constructed on prime of Alibaba’s open-source mannequin. 

An engineer overseeing ByteDance’s work on creating an agent, who spoke to MIT Know-how Assessment anonymously to keep away from sanction, stated that the absence of Claude Sonnet fashions “limits every thing we do in China.” DeepSeek’s open fashions, he added, nonetheless hallucinate too usually and lack coaching on actual‑world workflows. Builders we spoke with rank Alibaba’s Qwen collection as the most effective home various, but most say that switching to Qwen knocks efficiency down a notch.

Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human‑Centered AI, thinks that hole will shut: “Constructing agentic capabilities in base LLMs has change into a key focus for a lot of LLM builders, and as soon as folks notice the worth of this, it should solely be a matter of time.”

For now, Manus is doubling down on audiences it could already serve. In a written response, the corporate stated its “main focus is abroad enlargement,” noting that new workplaces in San Francisco, Singapore, and Tokyo have opened up to now month.

An excellent‑app strategy

Though the idea of AI brokers remains to be comparatively new, the consumer-facing AI app market in China is already crowded with main tech gamers. DeepSeek stays essentially the most broadly used, whereas ByteDance’s Doubao and Moonshot’s Kimi have additionally change into family names. Nevertheless, most of those apps are nonetheless optimized for chat and leisure fairly than activity execution. This hole within the native market has pushed China’s massive tech companies to roll out their very own user-facing brokers, although early variations stay uneven in high quality and tough across the edges. 

ByteDance is testing Coze Area, an AI agent based mostly by itself Doubao mannequin household that lets customers toggle between “plan” and “execute” modes, to allow them to both instantly information the agent’s actions or step again and watch it work autonomously. It connects as much as 14 common apps, together with GitHub, Notion, and the corporate’s personal Lark workplace suite. Early opinions say the software can really feel clunky and has a excessive failure price, but it surely clearly goals to match what Manus provides.

In the meantime, Zhipu AI has launched a free agent referred to as AutoGLM Rumination, constructed on its proprietary ChatGLM fashions. Shanghai‑based mostly Minimax has launched Minimax Agent. Each merchandise look virtually equivalent to Manus and demo primary duties resembling constructing a easy web site, planning a visit, making a small Flash recreation, or operating fast knowledge evaluation.

Regardless of the restricted usability of most normal AI brokers launched inside China, massive firms have plans to alter that. Throughout a Might 15 earnings name, Tencent president Liu Zhiping teased an agent that may weave automation instantly into China’s most ubiquitous app, WeChat. 

Thought-about the unique super-app, WeChat already handles messaging, cellular funds, information, and thousands and thousands of mini‑packages that act like embedded apps. These packages give Tencent, its developer, entry to knowledge from thousands and thousands of companies that pervade on a regular basis life in China, a bonus most rivals can solely envy.

Traditionally, China’s shopper web has splintered into competing walled gardens—share a Taobao hyperlink in WeChat and it resolves as plaintext, not a preview card. Not like the extra interoperable Western web, China’s tech giants have lengthy resisted integration with each other, selecting to wage platform battle on the expense of a seamless consumer expertise.

However the usage of mini‑packages has given WeChat unprecedented attain throughout companies that after resisted interoperability, from fitness center bookings to grocery orders. An agent in a position to roam that ecosystem might bypass the combination complications dogging impartial startups.

Alibaba, the e-commerce big behind the Qwen mannequin collection, has been a front-runner in China’s AI race however has been slower to launch consumer-facing merchandise. Regardless that Qwen was essentially the most downloaded open-source mannequin on Hugging Face in 2024, it didn’t energy a devoted chatbot app till early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search software. By June, Quark had launched DeepResearch—a brand new mode that marks its most agent-like effort thus far. 

ByteDance and Alibaba didn’t reply to MIT Know-how Assessment’s request for feedback.

“Traditionally, Chinese language tech merchandise are inclined to pursue the all-in-one, super-app strategy, and the most recent Chinese language AI brokers replicate simply that,” says Li of Simular, who beforehand labored at Google DeepMind on AI-enabled work automation. “In distinction, AI brokers within the US are extra centered on serving particular verticals.”

Pei, the researcher at Stanford, says that current tech giants might have an enormous benefit in bringing the imaginative and prescient of normal AI brokers to life—particularly these with built-in integration throughout companies. “The client-facing AI agent market remains to be very early, with tons of issues like authentication and legal responsibility,” he says. “However firms that already function throughout a variety of companies have a pure benefit in deploying brokers at scale.”