Zhongtai Securities released a research report saying that on March 6, 2025, the first general-purpose AI agent Manus came out, which can disassemble complex tasks into executable step chains, and flexibly call tools in a virtual environment to finally deliver complete results, which initially reflects the Agent's ability to work independently in the form, or marks the AI from Reasoner to Agent, and the Agent capability has entered the "Take Actions" stage. Vendors with more user traffic are expected to achieve a positive cycle of "traffic-data-user experience", and with the upgrade of open source model capabilities to make up for the technology generation gap between large and small manufacturers, the engineering capability of AI products may open up the gap in product experience.

The main views of Zhongtai Securities are as follows:

Agent = digital worker to achieve independent decision-making and task execution

AI Agent=LLM× (Planning + Memory + Tools + Action), which can perceive the environment, make decisions and execute actions independently, and have the ability to gradually complete goals through independent thinking and invoking tools, not only limited to benchmarking other forms of applications, but directly benchmarking manual labor. At present, TaskAgent is the mainstream, and the experience of multi-agent collaborative applications has been improved. The agent market has sufficient potential, helping the industry to reduce costs and reshape value, and there is a strong willingness in professional services, finance, healthcare, and retail scenarios, and the trillion-level Chinese AI market is expected to open.

The singularity is coming: the Manus universal agent is taking shape, ushering in a new era of AIAgent

The

focus of large model development has shifted to "post-training, inference side, and autonomous agent", and has transitioned from "volume base, volume model" to "volume inference, volume AIAgent, and volume application landing". On March 6, 2025, Manus, the first general-purpose AI agent, was launched, which can disassemble complex tasks into executable step chains, and flexibly invoke tools in a virtual environment to finally deliver complete results.

Industry outlook: Pay attention to the entrance, AI application + intelligent hardware ushered in industrial transformation

AI application: traffic is king, and product engineering capabilities are the follow-up focus. Vendors with more user traffic are expected to achieve a positive cycle of "traffic-data-user experience", and with the upgrade of open source model capabilities to make up for the technology generation gap between large and small manufacturers, the engineering capability of AI products may open up the gap in product experience.

1) General vs. vertical: General-purpose agents have higher requirements for generalization capabilities, platform-based companies have higher R&D endowments, large customer bases, wide product matrix, sufficient data accumulation, and abundant cash flow, which have greater advantages in the competition of general-purpose agents. Vertical agents need to strengthen their understanding capabilities in specific fields, and relevant vertical vendors are expected to differentiate themselves through customer resources, business understanding, and specific data accumulation.

　　2) ToCvsToB: The C-side agent focuses on the user experience of the product, and the essence is to return to the competition of traffic and payment rate. On the one hand, enterprises have a strong willingness to reduce costs and increase efficiency, and on the other hand, clear business scenarios, business logic, and industry data are more suitable for the use of agents.

Intelligent hardware: The device-side implementation of the agent is the last mile of the agent, and various types of intelligent hardware are blooming everywhere. With the acceleration of AIAgent technology iteration and industrialization process, intelligent hardware products may face opportunities for industrial upgrading and become the "traffic entrance" in the general or vertical fields.

Investment Advice:

Application: End-side cost revolution + autonomous intelligence, agent applications are expected to usher in an explosion.

DeepSeek Open Source Week has set off a cost revolution, the upgrade of model capabilities has accelerated the implementation of applications, and the intelligent upgrade of key link tool products has exploded in an all-round way. It is recommended to pay attention to: Straight Flush (300033.SZ), Kingsoft Office (688111.SH), INTSIG Information (688615.SH), EZVIZ Network (688475.SH), etc.

Computing power: The explosion of application demand corresponds to the growth of computing power consumption, and public cloud + private cloud + all-in-one machine are the three mainstream deployment forms, driving the domestic computing power core to benefit.

1) Public cloud: Cloud computing companies have increased CAPEX spending to promote the increase of computing power and effective computing power. It is recommended to pay attention to: Alibaba (09988), Tencent (00700), etc.

2) Private cloud: Privatization deployment is just needed by national conditions, and the willingness to purchase has increased. It is recommended to pay attention to: Haiguang Information (688041.SH), Sugon (603019.SH), Shensanda A (000032.SZ), etc.

3) All-in-one machine: Lightweight out-of-the-box for party, government and financial scenarios, the demand is hot. It is recommended to pay attention to: Digital China (000034.SZ), iSoftStone (301236.SZ), Fibersoft Communications (600498.SH), Topway Information (002261.SZ), Dahua (002236.SZ), etc.

Risk Warning:

The engineering capability of AI products is not as expected; Commercialization is not progressing as expected; The progress of model technology iteration is not as expected; There is a risk that publicly available information used in research reports may be delayed or not updated in a timely manner.