DeepSeek Harness Team Releases Super-agent Architecture

Article body

On the evening of October 31, DeepSeek released the Super-agent architecture, handing conversation orchestration over to the system and returning people to decision-making and acceptance.

Agents move from single conversations to parallel multitasking

On the evening of October 31, the DeepSeek Harness team published a technical article titled "Super-agent Design Exploration: Toward a More Autonomous Agent System," proposing to rebuild the Agent system around five major modules—a dual-layer architecture, event-driven design, asynchronous multi-channel interaction, separation of storage and compute, and an external Watchdog—and summarizing it into four characteristics: continuous, efficient, proactive, and persistent. The article argues that as models increasingly require less human intervention to execute a single task, the factor limiting productivity gains is shifting toward human attention, and Agents need to evolve from tools that passively respond to instructions into collaborators that are continuously online and can process multiple tasks in parallel.

The team also released preliminary experimental data. In synthetic load tests with about 300 messages each, of which about 20 were work requests, the median and 95th-percentile task completion times for Super-agent were 12.4 minutes and 15.7 minutes, respectively, while the 95th percentile for a single-conversation approach under neutral prompts was 29.5 minutes. In stress tests of the Router layer, as the event arrival rate rose from about 1 per minute to about 40 per minute, the inference time allocated per individual decision instead fell from 6.4 seconds to 1.6 seconds. An internal instance based on this approach has been running continuously, participating in issue reproduction and fixes and code review responses; when team members report software problems in group chats, even if no one @s the AI, it can identify defects and attempt to reproduce them in the background.

Policy groundwork has been equally intensive. The Ministry of Industry and Information Technology and seven other departments previously jointly issued the "Implementation Opinions on the 'AI + Manufacturing' Special Action," proposing to promote the deep application of 3 to 5 general-purpose large models in manufacturing by 2027, launch 1,000 high-level industrial intelligent agents, and promote 500 typical application scenarios. The previously issued "Opinions on Deeply Implementing the 'AI +' Action" clearly states that by 2027, the penetration rate of new-generation intelligent terminals, intelligent agents, and other applications will exceed 70%, and by 2030 it will exceed 90%.

Overseas resonance is reinforcing this expectation. Claude Sonnet 5.5, released by Anthropic on September 28, saw its score on Terminal-Bench 4.0 rise from 10.3% in the previous generation to 70.6%, and its OSWorld 2.1 score rise from 57.0% to 80.1%, with simultaneous leaps in multi-file programming and long-cycle knowledge work capabilities. Enterprise-side payment data is also following suit: Salesforce disclosed that its Agentforce-related annual recurring revenue exceeds $1.5 billion, up 240% year over year.

AI applications refer to the application-layer industry that embeds large-model capabilities into software products and business processes, extending "answering questions" into "completing tasks." Currently, Agents are beginning to take over cross-system, multi-step execution, and the industry's focus is shifting from model parameters to task delivery and commercial closed loops.

"In 2026, AI applications will officially enter a period of flourishing diversity." Sinolink Securities believes that models are moving from Chat toward long Agents capable of continuously executing tasks, with the scope of application expanding from content generation to completing business work; vendors with customer bases and industry experience are better positioned to turn capabilities into revenue, and picks-and-shovels stocks, Infra, customized IT delivery, and industry software are four main lines worth long-term tracking.

Estimates of market space are even more aggressive. The number of active Agents globally is expected to grow from about 28.6 million in 2025 to 2.216 billion in 2030, a compound annual growth rate of about 139%. Guosheng Securities, citing IDC data, points out that over the same period the number of tasks executed by Agents per year will climb from 44 billion to 415 trillion, with increasing task complexity driving a leap in Token consumption, and the center of gravity of the industry chain will tilt toward segments with edge intelligence and computing infrastructure layouts.

Against the backdrop of industrial value revaluation, Zhongtai Securities further points out that "the center of value in the AI industry is shifting from model invocation to task delivery and ecosystem services, and enterprise-level Agents have entered a commercialization window." The institution calculates that by September 2026, weekly Token usage by Agents had reached about 7 times that of ordinary human-computer interaction scenarios, of which nearly 90% came from cached input reuse; deeper invocation combined with lower caching costs provides cost support for the implementation of long-cycle complex tasks.

This judgment echoes the view of CSC. CSC believes that domestic Agents are expected to drive Token invocation and a new round of ARR growth by increasing per-user task density, and that the key variables in the next stage are effective task volume, usage retention, and paid conversion.

Performance of several concept stocks is expected to grow rapidly

According to the East Money concept sector, there are currently 308 stocks in the A-share market involved in the AI application concept, with a combined total market value of about 3.85 trillion yuan.

Hikvision ranks first with a size of 316.646 billion yuan, East Money ranks second with 278.471 billion yuan, and BOE A follows closely with 165.218 billion yuan; Hithink RoyalFlush, Kingsoft Office, iFlytek, Century Huatong, Focus Media, Aochuang Data, Supcon Technology, 360, and Sangfor Technologies, 9 stocks in total, each have market values exceeding 50 billion yuan.

In terms of future growth potential, according to consensus institutional forecasts, a total of 13 AI application concept stocks are expected to see profit growth of more than twofold this year. Among them, institutions expect Dian Diagnostics' net profit to surge nearly 11 times year over year, UTour Group's net profit to grow more than 7 times, and Maxvision Technology, AOTO Electronics, and two other stocks to see forecast net profit growth of more than 3.5 times; Yinli Media, ZWSOFT, Tianyu Digital Science and Technology, Aochuang Data, and three other stocks are also expected to see net profit growth of more than 2 times.

Supcon Technology has turned industrial AI into a subscribable service. In the first half of 2026, the company's industrial AI business revenue was 385 million yuan, up about 229% year over year; its TPT time-series large model has covered 13 industries including chemicals, petrochemicals, oil and gas, and pharmaceuticals, and has been implemented at customers such as PetroChina, Sinopec, and Wanhua Chemical. The company is advancing MaaS and RaaS subscription models; after Xingfa Group deployed 5 sets of TPT, the plant's staffing was reduced from 260 to 80.

Kingsoft Office launched both Lingxi Professional Edition for individuals and WPS Comate for organizations in July. In the third quarter, WPS 365 revenue was 270 million yuan, up 33.83% year over year; WPS Comate has been implemented in industries such as healthcare, energy, and education, and the monthly data analysis time of the nursing department at Peking University First Hospital was shortened from two to three weeks to two or three days. As of the end of September, WPS Office's global monthly active devices reached 678 million.

(Source: East Money Research Center) Middle of the body Embedded advertisement