Top AI Product

[Editor’s Note] 2025 is being widely hailed as the “Year of the AI Agent,” and China’s AI startup scene is experiencing an unprecedented “Agent craze” . From general-purpose players to those focusing on vertical scenarios, various companies are entering the fray, attempting to define the next era of AI. However, behind this booming technical race, a practical problem is increasingly evident: as the capabilities of Agents continue to expand, where are the truly high-value application scenarios that can support hefty R&D and operational costs and achieve scaled profitability? The pressure of ARR (Annual Recurring Revenue) is becoming a sword of Damocles hanging over the heads of many Agent companies.

If the past year was about the rapid advancement of large models, then entering 2025, the battle has shifted to the front lines of (implementation) – the AI Agent. These “digital laborers,” carrying high hopes, are no longer content with simple Q&A. Instead, they are designed to autonomously plan, use tools, execute complex tasks, and deliver results. Overnight, several highly representative Agent companies have emerged in the Chinese market, each bringing their technical philosophies and product forms to join this intensely competitive race.

A Panorama of Players: From General-Purpose and Search to Collaboration and Browser Agents

In this wave of Agent enthusiasm, the moves of several companies are particularly noteworthy. They are attempting to solve the challenge of Agent implementation from different angles:

Manus: The General-Purpose Agent’s “All-Around Warrior” Dream

Launched by Chinese startup Monica (also known as Butterfly Effect), Manus is a typical example of a general-purpose AI Agent. Its core vision is to achieve the automated execution of tasks “from idea to action.” Manus employs a multi-agent architecture to break down complex tasks and possesses multimodal capabilities, supporting web Browse, spreadsheet processing, code generation, and more, aiming to become a “super assistant” capable of autonomously completing travel planning, data analysis, content creation, and even resume screening.

Manus’s ambition lies in its generality, hoping to cover as many task scenarios as possible. However, generality often means needing to be sufficiently “good” in every vertical domain, which not only demands extremely high technical prowess but also faces challenges in user habit migration and trust building. The product is currently in internal testing, with invitation codes reportedly selling at high prices on secondary markets, reflecting market anticipation but also suggesting that widespread adoption will take time. Finding scenarios willing to pay a high premium for this general capability is key to its commercialization.

Genspark: The “Search” Narrative – Reshaping Information Retrieval with Agents

Founded by former Baidu executives Jing Kun and Zhu Kaihua, Genspark chose the more concrete entry point of “search.” Genspark’s core product is “Sparkpages,” generated in real-time based on a multi-agent framework. These customized pages integrate information from various sources and include built-in AI assistants. It attempts to overturn the traditional search engine’s link-list model with “high-quality, unbiased information” free of advertisements.

Genspark’s strength lies in addressing the user’s fundamental need for information retrieval and enhancing the efficiency of information integration and presentation through Agent capabilities. Focusing Agent capabilities on search and information organization is its strategy to differentiate itself from purely general-purpose Agents. However, the search domain is intensely competitive. How to build a differentiated user experience, cultivate user stickiness, and explore sustainable business models (high-value monetization methods beyond advertising) are questions Genspark needs to answer. Its significant funding indicates optimism from the capital market, but converting this into actual ARR requires continuous breakthroughs in user scale and commercialization pathways.

Coze Space: ByteDance’s “Platform” Play and Ecosystem Building

ByteDance’s Coze Space is positioned more as a collaborative platform for AI Agents. It aims to help users efficiently complete entire task workflows from requirement to outcome delivery through automated task breakdown, tool invocation, and result generation. Coze Space offers both exploration and planning modes and emphasizes MCP (Multi-Channel Plugin) extension capabilities, supporting user-defined plugins and building an ecosystem of expert Agents.

Coze Space adopts a typical platform strategy, aiming to attract developers and users by building an ecosystem. Its advantage lies in ByteDance’s strong technical foundation and user base. By providing tools and a platform, it lowers the barrier to Agent development and use, potentially rapidly accumulating B2B and B2C users. However, the success of a platform depends on the prosperity of its ecosystem. How to incentivize developers to contribute high-quality Agents, how to balance general capabilities with vertical needs, and how to achieve high-value monetization from platform services are directions Coze Space needs to continuously explore. Its B2B services, in particular, need to find enterprise-level application scenarios that can generate stable, high ARR.

Fellou: Practicing the “Action” Loop with a Browser Agent

Developed by a Chinese team of post-95s generation, Fellou deeply integrates Agent capabilities with the browser to create an “Agentic browser.” Fellou emphasizes “Deep Action” and “Proactive Intelligence,” capable of autonomously executing complex tasks across websites and applications within a sandbox environment, and possessing contextual memory and user behavior prediction capabilities.

Fellou’s innovation lies in upgrading the browser from an information portal to an action platform, attempting to realize the value of Agents within the digital environment users inhabit most frequently. By combining with workflow automation, Fellou directly targets scenarios that improve user efficiency and productivity. The task duration and cost data it provides attempt to demonstrate its efficiency in specific tasks. However, changing users’ long-standing browser habits is not easy. How to make users perceive the significant efficiency gains brought by Agent capabilities and be willing to pay for them is a hurdle Fellou’s commercialization must overcome. Compared to general-purpose Agents, the scenarios for a browser Agent might be more focused on information processing and process automation, requiring deep penetration in specific domains to find high-value payment points.

Calm Beneath the Frenzy: The ARR Predicament of General-Purpose Agents

From Manus’s general versatility to Genspark’s search innovation, Coze Space’s platform ecosystem, and Fellou’s browser action, these companies represent different directions in China’s AI Agent exploration. Their technical concepts and product forms are unique, collectively pushing the boundaries of Agent capabilities forward.

However, behind this intensely competitive trend, a reality that cannot be ignored is that most current general-purpose AI Agents still face severe commercialization challenges, particularly in finding high-paying application scenarios that can generate sustainable ARR.

Mismatch between Task Diversity and Willingness to Pay: General-purpose Agents aim to solve a wide range of problems, but many individual users or light users have a low willingness to pay for task automation, or are only willing to pay for very specific, high-frequency tasks. While enterprise users have the ability to pay, their needs are often highly customized and complex, which general-purpose Agents struggle to meet directly, requiring significant custom development and integration work that increases delivery costs and dilutes the attractiveness of ARR.

Efficiency and Stability in the “Last Mile”: Although Agents show potential in planning and partial execution, when dealing with complex, dynamic real-world tasks, the “last mile” implementation often encounters various issues, such as task failures due to website structure changes, understanding errors with unstructured information, and difficulties integrating with existing enterprise systems. These problems affect the efficiency and stability of Agents, leading enterprises to adopt a wait-and-see attitude when considering their adoption, making it difficult to form stable subscription payments.

ARR is the Core: For startup companies, consistent, predictable ARR is a key indicator of the health of their business model and future growth potential. Pure technical leadership or concept hyping cannot sustain long-term development. Agent companies must find scenarios that create significant, quantifiable value for users, for which users are willing to pay continuously. This may mean shifting from generality to deep cultivation in specific vertical domains, solving industry pain points, providing end-to-end solutions, thereby securing high-value customers and building healthy ARR streams.

All Roads Lead… to Vertical Scenarios?

In this “卷” of AI Agents, an increasing number of voices are suggesting that, given the current immaturity of general-purpose Agent capabilities and the unclear commercialization path, focusing on vertical Agents for specific industries may be more practical and commercially viable. Examples include intelligent investment research Agents for the financial sector, auxiliary diagnostic Agents for healthcare, and intelligent customer service and marketing Agents for e-commerce.

While these vertical Agents have a relatively limited scope of capabilities, they can understand industry knowledge more deeply, invoke industry-specific tools more precisely, and solve industry pain points more effectively. They are easier to integrate with existing business processes and data systems, bringing measurable efficiency improvements or cost savings to enterprises, thus being more easily converted into high-value ARR.

Of course, the exploration of general-purpose Agents remains important; they are the foundation for advancing Agent technology. But from the perspective of commercialization and ARR, the most pressing challenge for China’s AI Agent companies is how to leverage general capabilities to deeply cultivate and find the “work” that can truly generate business value and for which users are willing to pay continuously.

This wave of AI Agent continues. It is both a manifestation of technological progress and an inevitable result of market competition. Ultimately, the companies that stand out in this competition may not be the ones with the most “flashy” technology, but those that best understand user needs, best solve practical problems, and best build a healthy ARR model. After all, AI Agents are built for better “work,” and right now, one of their most important “jobs” is to prove their commercial value.