DeepSeek Harness Team Unveils Super-agent Architecture

① The new Super-agent architecture released by the DeepSeek Harness team reduces human involvement in low-level execution; users only need to communicate with the "AI leader" responsible for coordination, and it will assign tasks to the AIs under its command to advance separately. ② Building an Agentic butler-type assistant rather than a passive assistant is the future direction.

Science and Technology Innovation Board Daily, November 2 (Reporter Huang Xinyi) — The DeepSeek Harness team recently published a technical article on its official WeChat public account, "Super-agent Design Exploration: Toward a More Autonomous Agent System," officially unveiling its new Super-agent architecture for autonomous agent exploration. Its core goal is to resolve a fundamental contradiction: as AI capabilities grow stronger, humans' limited attention has become the new productivity bottleneck. This architecture reduces human involvement in low-level execution; users only need to communicate with the "AI leader" responsible for coordination, and it will assign tasks to the AIs under its command to advance separately.

DeepSeek stated that improvements in model capabilities have given rise to new application forms, and Personal Agents have begun to emerge in concentrated numbers. However, existing mainstream Agent designs are still unable to support this model at the scale of multitasking and team collaboration; the upper limit on multitasking depends on human attention.

Through approaches such as a two-layer architecture, event-driven design, and separation of storage and computation, the DeepSeek team has overcome these limitations. At the underlying level, the system takes over cumbersome session management and context synchronization, enabling humans to move further up from specific execution details to higher-level decision-making and acceptance.

DeepSeek calls this Agent form and its system architecture Super-agent. At present, an internal Super-agent instance built on this approach is already running continuously. When team members report software issues in a group chat, even if no one asks the AI for help, the system can autonomously identify defects, reproduce the problem in the background, and seek a fix.

Sun Zhenya, a research manager at IDC China, believes that DeepSeek's product and technical approach is consistent with the "butler-style agent" direction advocated by Muse, that is, a more autonomous Agent. At present, the interaction logic of most agents still follows the earlier GenAI approach, in which users manage conversations and sessions themselves. A better solution is task-centric routing and scheduling, or a task-management-centric model that helps users manage concurrent tasks and scheduled trigger tasks. In this direction, DeepSeek's Harness mainly addresses related issues through router-based routing.

Sun Zhenya stated that building an Agentic butler-type assistant rather than a passive assistant is the future direction.

"Build products that can proactively help users, combining users' memory, schedules, location, consumption records, and life rhythms to help users complete needs end to end, and based on users' habits and preferences, gain insight into users' potential needs. For example, a week before a business trip, remind them to book tickets, and directly provide flights and hotels that match the user's habits as references; when frequently purchased daily necessities are about to run out, place orders in advance and let the user only perform the final confirmation action, minimizing repeated information entry and operational interaction. From Muse to butler-type assistants, competition in personal assistant Agents is shifting from task-based delivery to a comprehensive contest of ecosystem, memory, and butler-style services," Sun Zhenya said.

From IDC's assessment results, mainstream agent products in the Chinese market currently already possess relevant capabilities. In terms of memory, most products can already well support various hierarchical and categorized memory systems, including user memory, and can retain user preferences, task context, and key needs. In terms of tasks, conventional end-to-end task completion capabilities are also quite good, and for complex tasks such as those requiring browser operations or desktop operations, some agents also perform well. In terms of ecosystem, some products have already moved beyond office scenarios and connected with life services ecosystems. For example, WorkBuddy has partnered with Meituan to help users claim coupons and initiate transactions, completing payment through WeChat Pay's AI-exclusive card. Other Agent products such as Baidu Dazi, Nami Work, Teleagent, and SenseTime Xiaohuanxiong also possess certain life service capabilities.

Sun Zhenya believes that the Chinese market is still dominated by super apps, with large amounts of services, data, and transactions enclosed within each platform. The experience of Doubao Phone also shows that it is difficult to succeed by bypassing platforms through technical means alone. Sun Zhenya suggests that the industry needs to establish open standards for agents, including how agent identities are identified, which services they can invoke, how transaction revenues are distributed, and who is responsible when problems arise, so that platforms, merchants, and agent vendors can all find reasons to participate and build an application ecosystem serving agents.

Cailian Press statement: The content of this article is for reference only and does not constitute investment advice. Investors who act on it do so at their own risk.

Business cooperation

Popular unlocks

[Trend Research Report · Company] Profit forecast raised by more than 100%, the company benefits from the high prosperity of AI servers, demand for high-value products continued to expand in the third quarter, and combined with the trend of new product volume ramp-up becoming established, it opens up long-term growth space; there is also a rare overseas-building-materials company whose third-quarter profit performance was outstanding

① Profit forecast raised by more than 100%; the company benefits from the strong boom in AI servers, demand for high-value products continued to expand in the third quarter, and together with the established trend of new product ramp-up, this opens up long-term growth space; ② This rare overseas-expanding building materials company delivered outstanding third-quarter profit performance, the high profitability of the overseas markets it entered is gradually being realized, and its M&A management projects are expected to see further incremental implementation soon; under globalization, its valuation framework is expected to be re-rated.

[Telegraph interpretation] ByteDance's Fanqie Novel will invest more than 1 billion yuan to support content creation in 2027. Institutions expect AI short dramas to see a dual explosion in scale and users in 2026, with the full-year market size expected to exceed 40 billion yuan. This company's short drama series have nearly 30 million likes on Douyin.

ByteDance's Fanqie Novel will invest more than 1 billion yuan to support content creation in 2027. Institutions expect AI short dramas to see a dual explosion in scale and users in 2026, with the full-year market size expected to exceed 40 billion yuan. This company's short drama series have more than 400 million video views and nearly 30 million likes on Douyin, while another company has self-produced a VR short drama.

Related reading

Space computing meets a reality check: Google energy chief says costs can reach 100 times those of ground data centers

Cailian Press, Zhou Ziyi, 2026-11-02 17:16

① Google's data center energy head said the launch cost of sending computing hardware into orbit is "extremely high, up to 100 times the cost of ground construction"; ② The real bottleneck lies in the approval process and lengthy assessment studies required to connect projects to the grid; ③ If launch costs fall to $200 per kilogram or below, the computing cost per unit of power in space is expected to be roughly comparable to ground levels.

Is the "space AI era" coming soon? Musk makes a bold prediction: orbital computing power will crush ground computing by 2029!

Cailian Press, Huang Junzhi, 2026-11-02 15:38

① Musk predicts that by 2029, AI and cloud computing processing capacity deployed in low Earth orbit will far exceed ground computing power; ② "Compared with orbital computing, ground computing power is simply negligible," he wrote.

Altman's Asia trip: just after meeting Samsung and SK executives, he heads to the Middle East to sign four agreements

Cailian Press, Shi Zhengcheng, 2026-11-02 15:31

① During a brief visit to South Korea, OpenAI founder Altman discussed AI transformation with about 50 Samsung executives and had lunch with SK Group Chairman Chey Tae-won, with attention on progress in supply chain cooperation; ② Subsequently, Abu Dhabi officials announced four agreements with OpenAI to bring AI into scenarios such as energy operations, government compliance review, and investment analysis, among which a food import review pilot has already been put into operation.