China is experiencing a surge in the development and adoption of AI agents, sophisticated systems designed to autonomously perform tasks for users. This trend follows the previous year’s focus on foundation models, which are the large language models that power much of the AI revolution. Now, the emphasis has shifted towards creating AI agents capable of handling a wide range of functions, from managing emails to planning complex itineraries.
The Emergence of AI Agent Startups
Numerous Chinese startups have recently emerged, dedicated to building general-purpose digital tools that can significantly enhance productivity and convenience. These AI agents are designed to assist with various tasks, including answering emails, conducting internet research for vacation planning, and even creating interactive websites. The rise of these startups has been rapid, with many appearing within the last few months, inspired by the success of Manus, an AI agent that generated considerable excitement on social media.
Unlike large language models (LLMs) that primarily respond to user queries, these AI agents are built upon them, employing a structured, workflow-based approach to accomplish tasks. This involves a different mode of interaction with AI, focusing on managing and executing multi-step processes such as booking flights, organizing schedules, and conducting in-depth research. What distinguishes these agents is their ability to leverage external tools and retain instructions, ensuring a seamless and efficient user experience.
China’s Potential Leadership in AI Agents
China is well-positioned to lead the development of AI agents, thanks to its unique ecosystem characterized by tightly integrated apps, fast product development cycles, and a digitally savvy user base. This environment is particularly conducive to embedding AI into everyday life, making it easier for users to adopt and integrate these tools into their daily routines.
Currently, leading AI agent startups in China are primarily targeting the international market. This is largely due to the fact that many of the most advanced Western AI models are inaccessible within China’s internet firewall. However, this situation is evolving as major tech companies like ByteDance and Tencent develop their own AI agents. These agents could be integrated directly into their existing super-apps, utilizing the vast data and resources within their ecosystems to enhance automation across various aspects of daily life.
The race to define the functionality and utility of AI agents is intensifying, with both ambitious startups and established tech giants exploring how these tools can be effectively implemented and for which user segments. This competitive landscape is driving innovation and pushing the boundaries of what AI agents can achieve.
Manus: A Trailblazer in the AI Agent Space
Manus, developed by the startup Butterfly Effect, has experienced remarkable growth. The company secured $75 million in funding led by the US venture capital firm Benchmark, embarked on a global roadshow to promote its product, and significantly expanded its workforce.
Even before its public launch in May, Manus had become a benchmark for what a versatile, consumer-focused AI agent should be capable of. Unlike AI agents designed for specific business tasks, Manus is designed to assist with a wide range of everyday tasks, such as planning trips, comparing stock options, and helping with school projects.
Manus offers a unique browser-based sandbox environment that allows users to monitor the agent’s activities in real time. This feature enables users to observe as the agent navigates web pages, reads articles, and codes actions. Additionally, Manus proactively seeks clarification from users and maintains long-term memory to provide context for future tasks, enhancing its ability to handle complex and ongoing projects.
According to Ang Li, cofounder and CEO of Simular, a startup focused on creating AI agents that operate real computers, Manus represents a significant advancement in AI agent user experience. He believes that Chinese startups have a distinct advantage in designing consumer products due to intense domestic competition, which drives rapid innovation and meticulous attention to product details. The combination of access to talent, venture capital, and a large domestic market are key drivers for this.
Competition and Global Ambitions
The competition in the AI agent market is rapidly intensifying. Emerging players like Genspark and Flowith are already demonstrating benchmark scores that rival or surpass those of Manus. These startups share Manus’s global ambition, with both Genspark and Flowith prioritizing the international market as their initial focus. The focus on the international market is more than just business strategy; it’s a necessity born out of the current limitations within the Chinese internet environment.
Chinese entrepreneur-led startups such as Manus, Genspark, and Flowith are increasingly poised to integrate into the global tech landscape and compete effectively on an international scale. Industry experts and investors suggest that these companies are characterized by their agility, efficient execution, and ability to rapidly develop innovative products. Their ability to adapt quickly to user feedback, iterate on designs, and incorporate cutting-edge technologies is a significant asset in this rapidly evolving field.
Financial incentives also drive the push towards launching overseas. Higher customer spending power and favorable exchange rates make international markets particularly attractive. As Manus cofounder Xiao Hong noted, pricing in USD can result in a significant revenue multiplier. This makes operating in international markets highly appealing, even if cultural differences may present some challenges. The ability to generate higher revenue streams allows them to reinvest in research and development and attract top talent. Understanding consumer behavior and preferences across different cultures is crucial for tailoring their products and marketing strategies.
Challenges and Opportunities in the Chinese Market
Creating similar AI agent functionality within China faces unique challenges. Major US AI companies like OpenAI and Anthropic have chosen to avoid mainland China due to geopolitical risks and regulatory complexities. This absence initially led to a black market where users relied on VPNs and third-party mirrors to access tools like ChatGPT and Claude. However, this gap has spurred the growth of local Chinese chatbots like DeepSeek, Doubao, and Kimi, although the demand for foreign models remains significant. This development has been encouraged and supported by the Chinese government, who sees AI as a key strategic industry.
Manus, for example, leverages Anthropic’s Claude Sonnet, which is widely regarded as a leading model for agentic tasks. Manus cofounder Zhang Tao has consistently praised Claude’s ability to manage tools, remember context, and engage in multi-round conversations, all of which are crucial for transforming AI from a mere chatbot into an effective executive assistant. The reliance on Claude Sonnet underlines the current dependence on best-in-class foreign models, even if they are not directly accessible within China.
However, the reliance on Sonnet means that Manus is effectively unusable within China without a VPN. Users accessing Manus from a mainland IP address will see a notification that the team is working to integrate Qwen’s model, a local version based on Alibaba’s open-source model. The availability of domestic alternatives is crucial for long-term growth within the Chinese market.
An engineer involved in ByteDance’s AI agent development stated that the lack of Claude Sonnet models significantly limits their capabilities in China. He noted that while DeepSeek’s open models are available, they still frequently hallucinate and are not adequately trained on real-world workflows. Developers generally consider Alibaba’s Qwen series to be the best domestic alternative, but switching to Qwen can result in a performance decrease. Balancing performance with accessibility is a constant challenge for those building AI agents within China.
Jiaxin Pei, a postdoctoral researcher at Stanford’s Institute for Human-Centered AI, is optimistic that these performance gaps will narrow over time. He believes that integrating agentic capabilities into base LLMs is becoming a primary focus for many LLM developers and that the value of these capabilities will drive rapid progress. He also expects improvements in areas like safety and interpretability, which are essential for wider adoption.
Focus on Overseas Expansion
Currently, Manus is prioritizing expansion in markets where it can readily serve users. The company has stated that its primary focus is international expansion, with new offices recently opened in San Francisco, Singapore, and Tokyo. These strategic locations provide them better access to talent, capital, and potential customers.
Though the concept of AI agents is still relatively new, the consumer-facing AI app market in China is already crowded with major tech players. DeepSeek remains the most widely used, while ByteDance’s Doubao and Moonshot’s Kimi have also gained popularity. However, most of these apps are still optimized for chat and entertainment rather than task execution. This gap has spurred China’s big tech firms to roll out their own user-facing agents which are currently uneven in quality and still considered to be rough around the edges. The focus on chat reflects the initial focus on developing large language models, but the trend is now shifting towards integrated AI agents.
Big Tech’s Response: ByteDance and Zhipu AI
ByteDance is currently testing Coze Space, an AI agent powered by its Doubao model family. Coze Space allows users to switch between “plan” and “execute” modes, enabling them to either directly guide the agent’s actions or allow it to work autonomously. It connects to 14 popular apps, including GitHub, Notion, and ByteDance’s Lark office suite. Early reviews suggest that while the tool can be clunky and has a high failure rate, it aims to match the functionality offered by Manus. The integration of various apps points toward the eventual goal of creating a truly interconnected and productive AI assistant.
Zhipu AI has launched AutoGLM Rumination, a free agent built on its ChatGLM models. Minimax, based in Shanghai, has launched Minimax Agent. Both of these products have a similar interface to Manus and demonstrate basic tasks such as building a simple website, planning trips, creating simple Flash games, or running data analysis. These demonstrations are aimed at providing a more tactile understanding of what AI agents can accomplish.
Tencent’s Super-App Strategy
Despite the limited usability of most general AI agents launched within China, major companies have plans to transform this landscape. During an earnings call on May 15, Tencent president Liu Zhiping hinted at an agent that would integrate automation directly into WeChat, China’s most ubiquitous app.
WeChat, considered the original super-app, combines messaging, mobile payments, news, and millions of mini-programs that function as embedded apps. These mini-programs provide Tencent, and their developers, access to data from a wide range of services, giving them significant advantages in providing integrated automation. The richness of data available within WeChat’s ecosystem gives Tencent a significant advantage.
The Interoperability Challenge
Historically, China’s consumer internet has been fragmented by competing walled gardens. For example, sharing a Taobao link in WeChat would only display as plaintext, without a preview card. Unlike the more interoperable Western internet, China’s tech giants have generally resisted integration with one another, prioritizing platform wars at the expense of a seamless user experience. This strategy, while beneficial for the companies involved, sometimes limited innovation and user experience.
However, mini-programs have given WeChat unprecedented reach across services that previously resisted interoperability, from gym bookings to grocery orders. An AI agent capable of navigating this ecosystem could bypass many of the integration challenges faced by independent startups. The mini-program ecosystem allows for a more standardized and integrated interface, which enables more efficient tasks execution.
Alibaba’s AI Efforts
Alibaba, the e-commerce giant behind the Qwen model series, has been a front-runner in China’s AI race but has been slower to release consumer-facing products. Despite Qwen being the most downloaded open-source model on Hugging Face in 2024, it didn’t power a dedicated chatbot app until early 2025. In March, Alibaba rebranded its cloud storage and search app Quark into an all-in-one AI search tool. By June, Quark had introduced DeepResearch, a new mode that marks its most agent-like effort to date.
The shift towards a search-based agent reflects the core business of the company and underlines their commitment for improved information retrieval and task automation by leveraging their huge resources.
The Future of AI Agents in China
According to Li of Simular, Chinese tech products have historically pursued an all-in-one, super-app approach, and the latest Chinese AI agents reflect this trend. He notes that AI agents in the US are more focused on serving specific verticals. The super-app approach is tailored towards ease of use and convenience, while the more specialized US model allows detailed customization and specialized problem-solving.
Pei, the researcher at Stanford, suggests that existing tech giants could have a significant advantage in bringing general-purpose AI agents to life, particularly those with built-in integration across services. He believes that the customer-facing AI agent market is still in its early stages, facing challenges such as authentication and liability. However, companies that already operate across a wide range of services are naturally better positioned to deploy agents at scale. The capability to access a wide range of users and their data, and the potential to quickly iterate on the first generation AI agents, and to quickly provide updates, are key determining factors in who are the leaders in this field.