Guotai Junan released a research report saying that recently, the Monica team released the world's first general-purpose AI agent product, Manus. The success of Manus demonstrates the potential of AI as a "digital agent", providing a new paradigm for AI to collaborate with humans. This model not only improves work efficiency, but also provides a new paradigm for AI to collaborate with humans, that is, AI can take on more complex tasks as an independent "agent". It has demonstrated strong cross-domain applicability and is expected to be deeply applied in more industries (such as healthcare, finance, education, enterprise management, etc.) in the future.

Manus has an innovative architecture operating model and powerful task handling capabilities.

Manus uses the Multiple Agent architecture, runs in an independent virtual machine, and greatly improves the processing efficiency of complex tasks through the division of labor and cooperation mechanism of planning agents, execution agents, and verification agents, and shortens response time through parallel computing. This architecture allows Manus to think and execute commands like a human, mimicking the way humans work, breaking down complex tasks into executable steps, and then invoking the right tools to complete the task. When it comes to task handling, Manus delivers the complete task directly, rather than providing advice or answers.

In addition, Manus' interaction design focuses on user experience, and synchronizes the progress of tasks in real time, so that users can clearly see the execution process of tasks, enhancing the user's sense of control. Manus also has the ability to memorize, which can optimize the output form of subsequent tasks according to the user's preferences, which improves the user's satisfaction.

In the GAIA benchmark, Manus achieved a new state-of-the-art (SOTA) performance on all three difficulty levels, surpassing OpenAI's large models at the same level.

The GAIA benchmark is an authoritative benchmark for evaluating the ability of general-purpose AI assistants to solve real-world problems. Proposed in 2023 by research teams such as Meta AI and Hugging Face, the test contains 466 questions divided into three difficulty levels, and aims to measure the ability of AI systems in various aspects such as inference, multimodal processing, web browsing, and tool calling.

In the GAIA benchmark, Manus achieved SOTA results, surpassing OpenAI's large models at the same level. It shows excellent robustness in solving complex tasks and long-tail problems. This result shows that Manus not only has strong task planning and execution capabilities, but also can flexibly call multiple tools to complete tasks through self-learning and cross-domain collaboration. Manus' success is largely due to the Multiple Agent architecture, which enables efficient disassembly and parallel processing of tasks by planning, executing, and validating agents working together. In addition, Manus's test configuration is consistent with its production version, ensuring reproducibility of results.

Recommended end-side SoCs: Zhongke Lanxun (688332.SH), Jingchen (688099.SH), Rockchip (603893.SH). Related benefits: Espressif Systems (688018.SH), Allwinner Technology (300458.SZ).

Catalyst: The promotion of AI applications has exceeded expectations.

Risk Warning: The actual application effect and scenario of Manus are not as expected.