OWL Team's In-Depth Share: The Best Team at Reproducing Manus, How Do They View the Current State of Agentic AI Adoption?

Manus went viral, making Agent one of the most watched AI areas in the first half of 2025, and also bringing developer attention to open-source projects that replicate Manus. OWL, launched by the CAMEL-AI team within 1 day after Manus went live, is one representative: the project's measured GAIA performance reached 58.18%, surpassing the 55.15% achieved by Huggingface's Open Deep Research ^[1].

In early March, Founder Park invited the OWL team to hold an online closed-door sharing, discussing the OWL technical framework, Manus, as well as Agent technical principles, implementation logic, and current commercial adoption. CAMEL-AI founder Li Guohao said that the OWL project and Manus are not completely the same, and there are many technical differences, but what they do is similar ^[1].

A key judgment from the sharing was: Manus's emergence showed the public the possibilities of AI technology, especially the current practical applications of agents, igniting this wave of AI agent technology; reproducing Manus is relatively simple technically, and lies more in product interaction and form, and Manus has first-mover advantage, so it will be relatively difficult for later products to reproduce its success ^[1].

On the tool ecosystem, the OWL team believes MCP is the future: it can let all frameworks access the same tools, just as Cursor and OWL can both use tools that comply with the MCP standard, and leverage numerous open-source tools to improve agents ^[1]. Regarding Agentic AI, the team also proposed that foundation models + external engineering frameworks are not the future trend; if work in a vertical domain can be easily replaced by a general agent, then it shows that work in that vertical domain is not 'vertical' enough and has not solved the most core pain point of that domain ^[1].

Another closed-door discussion in Silicon Valley, organized by Global Ready and Integer Intelligence, focused on Agent technical bottlenecks and commercialization. Guests pointed out that Manus, released in March, let the public feel the power of Agents for the first time; R1, o3 and more reasoning models provided a sufficiently strong technical foundation for Agent development, and 2025 is truly the year of Agents; however, Agents still face technical bottlenecks such as greater tool-use ability and longer context, and in the future there will also be how Agents collaborate with one another ^[7].

The adoption of reasoning models is also accelerating. After DeepSeek R1 launched, Volcano Engine was one of the fastest cloud platforms to deploy R1. Sharing organized by Volcano Engine and Founder Park and others noted that enterprises are paying more attention to how to combine DeepSeek with their own business scenarios, how to integrate the model for development, and how to mix generative models with reasoning models to achieve better results ^[5].