At a tech summit in San Francisco in late June, attendees at several closed-door sessions found a simple white A4 flyer placed on every seat. Printed in bold black text was a message that turned heads: “USD 36 million ARR (annual recurring revenue) in 45 days, eight products in ten weeks.”
The numbers referred to Genspark, a general-purpose artificial intelligence agent developed by MainFunc, a company founded by Jing Kun, former vice president at Baidu. Jing was formerly the CEO of subsidiary Xiaodu.
“Companies in the US were stunned. Everyone was asking what the company behind Genspark was,” an AI professional based in Silicon Valley told 36Kr.
Since the beginning of 2025, interest in Chinese-built AI agents has surged within the global tech community.
At that same summit, Genspark was mentioned repeatedly, alongside Manus, one of the first AI agents to gain widespread attention. “The last time Chinese AI got this much buzz was when DeepSeek launched its V3 and R1 models,” the source said.
It wasn’t just investors taking notice. Tech figures like Elon Musk also weighed in on social media, engaging with posts about Chinese-developed agents like Manus and Lovart.
The hype wasn’t accidental. It began with the rise of Manus, often described as the AI agent that opened the floodgates.
Created by the Monica team, Manus quickly gained traction. Within its first month, it reached 23 million monthly active users. One month later, it raised USD 75 million in a Series A round led by Benchmark, pushing its post-money valuation above USD 500 million. Against a backdrop of rising geopolitical tension, this marked a rare overseas success for a Chinese-founded startup.
What is an AI agent?
AI agents are applications that can independently plan tasks, invoke external tools, and deliver results. Unlike simple chatbots, these agents are designed to execute real-world tasks, essentially acting as digital assistants with operational capabilities.
The concept has been around. In June 2023, Lilian Weng, then head of applied research at OpenAI, published a 6,000-word blog post on autonomous agents powered by large language models. The post galvanized early interest in the space.
But at the time, agent capabilities were limited. Most were constrained by the capabilities of their underlying models and resembled glorified chatbots or robotic process automation (RPA) tools with limited utility.
When OpenAI launched GPT Builder in November 2023, backlash followed. Critics argued it lowered the barrier to entry too far, making it difficult for startups to maintain any competitive edge. Interest faded quickly.
“I invested in a few agent startups back in 2023,” one investor told 36Kr. “By 2024, they were struggling. Most of them pivoted back to traditional RPA, doing painful custom builds.”
The 2025 renaissance
AI agents are back in 2025, this time with stronger technical underpinnings, clearer product formats, and growing user and revenue numbers.
Two developments sparked the revival: Claude and Manus.
Claude, a model developed by Anthropic, has received significant upgrades this year. In February, it launched Claude 3.7 Sonnet, a hybrid reasoning model with improved programming abilities. In November 2024, Anthropic introduced the Model Context Protocol (MCP), which enables agents to invoke external tools without custom integrations.
Meanwhile, Manus offered more than just marketing. It provided developers with a workable product blueprint.
Manus’s interface, combining a dialogue window with a visual execution panel, became the standard for emerging applications like Lovart, YouWare, and Fellou.
A person close to Lovart’s developer, LiblibAI, said the company had long wanted to build a design-oriented agent but hadn’t settled on a format. “We started aggressively building three days after Manus launched,” they said.
Cracks behind the boom
Despite the momentum, signs of strain have appeared.
Manus’s monthly visits dropped from 23.76 million in March to 17.30 million in June, according to Similarweb. Genspark saw a smaller, but still notable, decline from 8.88 million in April to 7.69 million in June.
Geopolitical tensions have added uncertainty. In June, Manus laid off staff in China, scrubbed its presence from Chinese social media, and assembled a new team in Singapore. These moves were framed as steps toward operational efficiency and international expansion, but many observers saw them as a response to mounting scrutiny, especially from US regulators concerned about the company’s Chinese backers.
“In the future, any company hoping to raise money overseas will likely have to pick a side,” one investor said. “It’ll be a binary choice between markets.”
Within three months, the AI agent boom had shifted from explosive growth to existential questions.
One attendee at the “VB Transform” event described running into Genspark’s Jing Kun. “He looked visibly anxious,” the attendee said. “At one point, I overheard him ask: what do we do next?”
RELATED ARTICLE
Breaking the mold
Two patterns have emerged in the 2025 AI agent surge.
First: agents are hitting revenue milestones at breakneck speed. Genspark reportedly reached USD 10 million in ARR in just nine days. By comparison, Cursor, arguably the current benchmark for AI-assisted coding, took 21 months to do the same.
While the exact accounting for ARR may be imprecise, the signal is clear: after two years of market education, users are more willing than ever to pay for AI-powered tools. That shift has created fertile ground for monetization.
Second: the general-purpose agent category, initially the domain of large tech firms, has seen its breakout hits come from startups.
Data compiled by 36Kr shows that since early this year, startups have been leading in product launches, monetization progress, and influence.
Baidu’s agent Xinxiang, for instance, remains free to use but has low daily active users. A source familiar with the project said the team has already scaled back investment just three months after launch.
While larger companies are typically better at scaling generalized products, they also face institutional constraints. One Baidu employee noted that Xinxiang, which targets China’s market, cannot integrate overseas models like Claude or Gemini due to policy restrictions. “Engineering around that has been exhausting,” they said.
Startups, by contrast, have more flexibility. Agility has become their biggest advantage.
Generative AI researcher Zeng Yiqin summed it up well on X. “Manus is a good product,” Zeng wrote. But he followed on by claiming that Manus does not represent a “technology breakthrough” because it mostly follows a predefined flow to execute tasks, and works in a “relatively constrained environment.” That’s emblematic of today’s leading AI agents: responsive builds layered on third-party models, tailored for product-market fit.
Speed and specificity
The Manus team had just 50 people and brought the product to market in three months. Lovart’s entire development cycle was even shorter: less than two months.
The window of opportunity is narrow. Among the wave of Manus-inspired clones claiming to be pioneers in agentic AI, only Lovart has managed to sustain real momentum.
But speed alone isn’t enough. Founders need to clearly define use cases, refine product structure, and improve user experience through thoughtful engineering.
Even when built on external models, well-designed wrappers still matter. “The shell has its value,” said Monica founder Xiao Hong—a phrase that has since become common shorthand in AI circles.
In a blog post published on July 18, Manus co-founder and chief scientist Ji Yichao (also known as Peak Ji) revealed that the team had already restructured its agent framework four times to enhance performance and scalability.
Is the real future in vertical use cases?
But how useful are general-purpose agents, really?
Even for relatively simple tasks like website generation or equity report parsing, users have reported that Manus still struggles with hallucinations, poor intent recognition, and inconsistent outputs. The polished demos often fail to reflect day-to-day performance.
Usage trends support that skepticism. Traffic for both Manus and Genspark has declined since their respective launches, suggesting that neither has yet achieved a stable product–market fit.
“As a paying user, I’d rather just use Claude 4,” one developer said. “The quality is better.”
This highlights a key vulnerability for startups in the general-purpose agent space: their heavy reliance on external models.
“Eventually, OpenAI or Anthropic will own general-purpose agents,” one investor told 36Kr. “For companies that control the best models, making AI agents is the easy part.”
Case in point: OpenAI’s own agent, “Deep Research,” is already available to ChatGPT Pro subscribers. Around 20% of users have upgraded to the USD 200 monthly plan just to access it.
As for the many Manus-inspired spinouts, the same investor was blunt: “They will pivot to vertical use cases.”
That shift is already underway. Several companies originally focused on specific verticals are now integrating agent-like tools into their offerings.
LiblibAI, which began as an image model community in 2023, now operates Lovart, a design-focused agent. Cai Haoyu, a co-founder of Mihoyo, recently launched Whispers from the Star, an AI-powered game where each character functions as an agent capable of dynamic storytelling through user interaction.
Model performance still falls short of supporting truly universal AI agents. But in vertical domains, where the context is more constrained and the needs are clearer, agents are starting to deliver meaningful value.
In these settings, domain knowledge and networked resources form the most defensible moats. They shape product design and determine how well agentic AI can be embedded into professional workflows.
For startups, this approach also provides a strategic buffer, offering a more sustainable path forward as big tech continues its advance.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.