Chinese Start-Up Unveils World’s 1st Autonomous Agent

Chinese start-up Monica has launched Manus, the world’s first “AI Agent” platform — a significant leap beyond traditional large language models (LLMs) or chatbots. Unlike conventional AI tools, Manus functions autonomously, by initiating tasks on its own, capable of planning and executing complex, multi-step tasks across a wide range of domains. It has already demonstrated abilities such as building and deploying websites, analysing stocks, and generating detailed reports with minimal human involvement. This breakthrough has led many to call it the “second DeepSeek moment,” suggesting that China could be rapidly closing the gap in AI and potentially challenging the dominance of major US AI companies. Monica is reportedly preparing for a new funding round that could boost its valuation to $5 billion—a sharp rise from $1 billion at the end of last year. Alongside its fundraising efforts, Manus has also launched its monetization model, offering two subscription tiers: (1) Manus Starter: $39/month and (2) Manus Pro: $199/month. The difference is in the number of tasks that it can run simultaneously.

The secret recipe behind Manus’ success is that the company leverages a multi-model architecture that blends the capabilities of several leading language models. In practice, this means Manus builds its agents on top of third-party foundation models, such as Anthropic’s Claude 3.7 and fine-tuned versions of Alibaba’s open-source Qwen. However, this approach introduces two key challenges: (1) High Operational Costs: Manus reportedly pays $2 to Anthropic per completed task, which could significantly increase operational expenses as user activity scales. This cost burden is likely one reason Manus has adopted an invite-only system, resulting in just 186,000 registered users—about 1% of its total waitlist. (2) Limited Moat: Relying heavily on third-party and open-source technologies may reduce Manus’s defensibility, as competitors could potentially replicate its product by combining similar tools and models. A recent MIT Technology Review also pointed out technical limitations. The platform occasionally displays error messages indicating infrastructure bottlenecks. Additionally, the AI sometimes misinterprets user prompts, makes incorrect assumptions, or cuts corners to speed up task execution. Those issues may be addressed with its funding round.