文 | 财能圈
10月31日晚，DeepSeek Harness团队在官方公众号发表技术文章《Super-agent设计探索：迈向更自主的Agent系统》，配套原型dsh-superagent-prototype同日在GitHub开源，MIT协议。
文章开头有一句自我限定：这是一篇面向工程实践的架构思路探讨，并非对现有产品功能的介绍。
但文中给出的判断很具体：随着模型能力提升，AI完成单个任务已越来越不需要人工干预，人管理多任务的注意力却没有同步增长，一个人依然只能同时跟进3到5个活跃会话。DeepSeek的原话是：单任务的上限取决于模型能力，多任务的上限取决于人的注意力。
这句话是整篇文章的起点。文章还提到，模型能力的提升催生了一批新形态，个人助手开始集中涌现，它们直接接入消息、日程、邮件与工作平台，持续观察信息流。这种持续在线、事件驱动的工作方式不局限于个人场景，在软件开发等商业场景同样成立，但现有主流智能体的设计撑不住这个规模。
新架构的所有设计，都在解决这个卡点。
01 会话之间不通，上限就落在人身上
据悉，现有智能体以单个会话为基本单位，会话之间彼此独立。这会导致一种现象，A会话里得出的结论，B会话里的它并不知道。当用户并行推进多项工作时，必须手动切换，自己记住每一条上下文。到最后，只有人知道所有任务的全貌。
DeepSeek的解法是给AI分工。一层叫router，是持续运行的全局协调节点，统一接收群聊、GitHub、邮件等多渠道消息，维护共享上下文，把任务分派下去，自己不深入执行。另一层叫worker，各自维护独立上下文，并行做问题复现、代码审查、功能开发，必要时可调用Subagent（子智能体）。
这个设计和常见的“主智能体带几个子智能体”不一样。主智能体需要理解任务细节、规划子任务、高频协调，推理负担很重；router只管这件事该推给谁，worker之间上下文隔离，冲突靠订阅同一个话题自行发现。worker接手任务后还能直接订阅后续消息，省掉经router转发的协调开销。
DeepSeek自己说，这套设计没有提出全新理论，是把计算机系统里的经典工程思想搬到了智能体身上。
事件驱动，让群聊消息、邮件、日程提醒、GitHub评论、CI测试结果都成为触发行动的事件。存算分离，执行循环放在云端，会话状态存入数据库，本地只留一个轻量daemon（守护进程），电脑离线任务也能继续跑。外部watchdog（看门狗）负责检查任务是否陷入无效循环并尝试修复。异步ReActLoop（推理—行动循环）让AI执行耗时任务时仍能响应新指令，这一项尚未默认开启。
另外，还有一层投机执行，当结果是否会被采用尚不确定时，先在后台提前尝试，有价值再提交，没有就放弃。
此外，“主动”是这套系统的关键词，也是最需要克制的地方。绝大多数群聊消息不需要AI介入，DeepSeek设了两级过滤：先用程序规则筛掉明显无关的消息，再由大模型判断剩下的是随口吐槽还是真有问题，两道都通过才派任务。
文章给了一个内部实例：团队成员在群聊里说“这个项目有BUG”，没有人@它，Super-agent在后台静默复现，确认问题存在并找到简单修复方法后才提交PR（拉取请求），交给人验收。后续有人在GitHub评论里要求换一种修法，它又能从聊天软件的上下文切到GitHub继续做。这套策略在内部叫“静默复现并修复”。
文章还给了一组实验数据，基于该思路的Super-agent，任务完成时间中位数12.4分钟，95分位15.7分钟，优于带子工具的普通会话。DeepSeek同时说明，这是基于合成负载与简化环境的初步实验，不是正式产品结论。
产品边界也写在明处。GitHub上的原型是研究原型，不是正式产品。按仓库说明，公开发布的版本与DeepSeek内部运行的系统并不相同，文中部分功能仍在开发中。目前它主要面向一个GitHub仓库，也可以选择接入飞书。Super-agent只在被请求、或团队事先写好规则允许时行动，没有结果时不播报过程，没有权限时不擅自行动。
值得注意的是做这件事的团队。发布者是Harness团队，不是模型团队。
02 两轮融资之后，对外输出的东西变了
把时间线拉出来，这条新闻的位置就清楚了。
2026年4月，DeepSeek启动成立以来的首轮外部融资，6月底完成交割，规模约510亿元，投后估值接近4000亿元。出资方中，梁文锋个人出资约200亿元，腾讯约100亿元，宁德时代体系约50亿元，网易、京东、砺思资本、IDG资本各约30亿元，国家人工智能产业投资基金也在其列。
10月22日起，新一轮融资的消息在市场间传开：拟筹约1000亿元，最初募资目标500亿元，此后至少两次上调。腾讯和宁德时代是本轮出资最多的投资方。多家媒体从不同信源交叉求证，截至发稿，DeepSeek及相关投资方均未公开正式回应，最终金额仍未确定。
两轮相加，半年内累计融资规模约1500亿元。
钱有去处。据公开披露，7月它参与长鑫科技打新，8月斥资1.41亿元战略配售宇树科技，10月计划领投一家机器人初创企业；更早的报道显示，它还在内蒙古乌兰察布推进部署至少16万颗华为昇腾加速芯片的数据中心。
上市也在同步推进。自8月下旬起，DeepSeek接触多家投行，传闻中信证券已经进场尽调，市场说法是最快2026年末提交上市申请、2027年挂牌；9月下旬，高瓴创投原合伙人严文韬加入出任CFO。21财经9月的报道显示，辅导协议尚未签署，2027年的时间表属于市场传闻。这些说法均未获官方证实。
把这些和Harness放在一起看，指向同一件事：DeepSeek对外的标签一直是模型，V3、R1这些开源模型定义了它的身份。融资之后，除了继续训模型、买算力，它开始输出“怎么干活”；从过去只交付一个更先进的模型，转向交付一套替人调度、执行、验收的系统。
03 入口已经挤满，DeepSeek选了开源
值得注意的是，Super-agent选在10月底发布，同一个窗口期，这条赛道已经站满了人。
新浪财经在近日的报道里给了一个概括：智能体的竞争，正从执行具体任务，走向持续提供个人服务，成为真正的个人代理。
同期在做同一件事的名单不短。字节跳动8月25日发布面向生产力场景的智能体品牌“豆包工作”，与飞书深度打通，用户授权后能跨软件操作电脑；腾讯的Workbuddy早已上线试运营；阿里千问宣布加速打造Personal Agent（个人代理）；蚂蚁的灵光开启新版本内测。
月之暗面则走的是另一条路，一边把Kimi迭代到3万亿参数、首次推出Pro版本，一边在9月启动“Kimi企业合作伙伴计划”，拉上华胜天成、金山云等IT服务商，帮助企业把AI部署进核心业务。
企业侧的动作也在加密。腾讯在10月29日称，其智能体开发平台ADP可让企业定制、管理、协调用于客服、营销和运营的智能体；阿里云7月已推出Agent Loop与Agent Teams，配合原有的Agent Run，做智能体的开发、部署与监控。
争的是同一个位置：AI的下一个入口。聊天是入口，办公是入口，更可能被验证为入口的，是那个不用人开口就把事情办了的智能体。
DeepSeek入场的姿势和前几家不一样。它把底层Harness开源，模型、工具、技能、会话、沙箱，一切皆插件，MIT协议；Super-agent原型同样开源，并且从自己的研发场景切入，第一个场景是常驻代码仓库的Coding Agent。
这个姿势对应用层有直接影响。过去应用层的护城河是“更懂怎么调用模型”，一旦底层模型公司把Agent框架和最佳实践都开源，这层能力差会迅速变薄。对DeepSeek自己也是新战场：以前只跟OpenAI、Anthropic比模型，现在要跟字节、腾讯、阿里比产品、比场景、比谁能把智能体实实在在地放入真实的工作流。
文章里有一句是DeepSeek对自己的定位：“系统在底层接管了繁琐的会话管理与上下文同步，使人类得以从具体的执行细节，进一步上移到更高层的决策与验收。”
这是公司对自己产品的定义，不是对行业的判断。要检验它，未来得看两件事：内部那个实例跑了多久、跑成了什么；公开发布的原型能不能从“一个GitHub仓库”扩到更多场景。
最后，文章没有披露那个内部实例的运行规模，也没有给出任务成功率。公开的部分是一个研究原型、一个GitHub仓库和一组基于合成负载的实验数据。因此，一切还有待检验。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, the financial circle.
On the evening of 31 October, the DeepSeek Harness team published a technical article entitled Super-agent Design Exploration: Towards a More Autonomous Agent System in official WeChat Official Account, along with the prototype dsh-suberagent-prototype at GitHub Open Source, MIT protocol.
At the beginning of the article, there was a self-defining statement: it was a conceptual exercise oriented towards engineering practice, not an introduction to existing product functions.
However, the judgement given is very specific: as model capabilities have improved, AI has become less and less requiring manual intervention to perform individual tasks, attention to managing multiple tasks has not increased simultaneously, and one person can still follow only three to five active sessions simultaneously. DeepSeek is saying that the upper limit for a single mission depends on model capabilities and the upper limit for a multi-task depends on human attention.
This is the starting point of the whole article. The article also mentioned that the upgrading of modelling capacity had created a new pattern in which personal assistants had begun to come together and had direct access to news, schedules, mail and work platforms to observe information flows on a continuous basis. This sustained online, event-driven working approach is not limited to individual scenarios, but is equally valid in business contexts such as software development, but the current mainstream AI agent design cannot sustain this scale.
All the designs for the new architecture are working on this card point.
If there's no connection between sessions, the ceiling falls on people.
It is understood that the existing AI agent is based on individual sessions, which are independent of each other. This could lead to a phenomenon in which the conclusions reached in the A session were not known in the B session. When users advance multiple tasks in parallel, they must switch manually and remember each context. In the end, only one knows the full picture of all the missions.
DeepSeek’s solution is to divide the work of AI. The first floor, called Router, is an ongoing global coordination node that integrates information from multiple sources, such as community chats, GitHub and mail, maintains the sharing of context, assigns tasks and does not carry them out in depth. The other floor is called the worker, each maintaining its own independent context, with parallel problem recurrence, code review, functional development, and, if necessary, Subagent (sub-AI agent).
This design and common "Major AI agent with several AI agent" are different. Main AI agent needs to understand the details of the mission, plan subtask, and HF coordination, with a heavy burden of reasoning; Router only cares to whom the matter is to be pushed, and context is isolated between workingr, and conflicts are detected by subscription to the same subject. When you take over, you can also subscribe directly to follow-up messages and save the coordinated costs forwarded via router.
DeepSeek himself stated that the design did not introduce a whole new theory and that it moved the classic engineering thinking from the computer system to AI agent.
Events are driven by the fact that the news, mail, calendar alerts, GitHub comments, and the CI test results are all trigger events. The storage separation, the execution cycle is in the cloud, the session status is stored in the database, only one light daemon (dæmon) is left locally, and the computer offline mission can continue to run. External watchdogs are responsible for checking if the mission is in an ineffective cycle and trying to repair it. The rect Loop (inference-action cycle) allows AI to respond to new instructions while on a time-consuming mission, which has not yet begun by default.
In addition, there is a layer of speculation that, when it is uncertain whether the results will be used, it will be attempted in advance and valuable before submission, without giving up.
In addition, “active” is the key word for the system and is where restraint is most needed. The vast majority of groups talk about information without the involvement of AI, and DeepSeek has set up two-tier filters: first, using the rules of procedure to sift out apparently unrelated news, then LLM to judge whether the rest is random or really problematic, both of which go through the task.
The article gave an internal example: team members said in the group conversations that “the project has BUG”, no one @, that Super-agent returned silently in the background, confirmed the problem and found a simple repair method before submitting the PR (raised request) to person acceptance check. Later, in GitHub's review, a request was made for a change of law, which in turn could go from context to GitHub in chat software. The strategy is internally called "quiet relapse and repair."
The article also gave a set of experimental data based on which Super-agent had a median of 12.4 minutes, 95 minutes of 15.7 minutes, which was better than the normal session of the tape tool. DeepSeek also stated that it was a preliminary experiment based on synthetic load and simplified environment and not a formal product conclusion.
The product boundary is also stated. The prototype on GitHub is a research prototype and not a formal product. According to the warehouse, the public version of publish is not the same as the system operated internally in DeepSeek, and parts of the text are still under development. Currently, it is mainly aimed at a GitHub warehouse, with the option of accessing the Flying Book. Super-agent acted only when requested or when the team had written the rules in advance, and did not broadcast the process when there was no result, and was not authorized to act without the permission of limited-time.
It's worth the team doing this. publish is a Harness team, not a model team.
After two rounds of financing, something from abroad changed.
Pull out the timeline and the news will be clear.
In April 2026, the first round of external financing since DeepSeek was launched was completed at the end of June, with a size of approximately $51 billion and a post-investment valuation of nearly $400 billion. Among the donors, Liang Wenbong personally contributed about 20 billion yuan, Tencent about 10 billion yuan, the Ningde era system about 5 billion yuan, the Internet, Kyoto, Guangxi and IDG capital respectively, and the National Investment Fund for Artificial Intelligence.
Since 22 October, news of the new round of financing has spread across the market: about $100 billion is proposed, with an initial fundraising target of $50 billion, and at least twice thereafter. The era of Tencent and Ninde was the largest investor in the current round. Many media cross-checked from different sources, and as of the time of distribution, DeepSeek and the investors concerned had not responded publicly formally and the final amount remained undetermined.
Together, the cumulative amount of financing over a six-month period is approximately $150 billion.
Money goes to places. According to public disclosure, in July, it participated in a new Changsung technology campaign, in August it spent $141 million on strategic tree distribution technology, and in October it plans to take over a robot-starter; earlier reports indicate that it is also pushing for the deployment of at least 160,000 data centres in Ulanchabu, Mongolia, to boost the chip.
The market is also advancing simultaneously. Since late August, DeepSeek has been in contact with a number of investment banks, and there are rumours that the CFO was joined in late September with the market saying that it was the fastest time to submit a listing application in late 2026 and that it would be closed in 2027. Financial reports in September indicate that the mentoring agreement has not yet been signed and that the timetable for 2027 is market hearsay. None of these claims have been officially substantiated.
Put these together with Harness and point to the same thing: DeepSeek's external label has been a model and the open source models V3, R1 define its identity. After financing, it began to export “how to work” in addition to continuing training models and purchasing power, moving from delivering only one more advanced model in the past to delivering a system of alternative movement control, execution, acceptance check.
The entrance is full. DeepSeek picked the open source.
It is worth noting that Super-agent was chosen at the end of October for publish, the same window period when the track was already full of people.
In recent reports, the New Wave has given an overview: the competition for AI agent is moving from carrying out specific tasks to the continuous provision of personal services to becoming a real personal agent.
The list of persons doing the same thing during the same period is not short. On August 25, byte beats the productivity scene of AI agent's “Doubao Job”, which connects to the depths of flying books and allows users to cross software to operate their computers; Tencent's Workbuddy is already on line for piloting; Ali is asking for an announcement to speed up the construction of Personal Agent (personal agent); and the ants' luminous light opens new versions of internal measurements.
The dark side of the moon is another path, one that replaces Kimi with 3 trillion parameters, the first release of the Pro version, and the launch of the Kimi Business Partnership in September, bringing together IT service providers such as China Victory and Jinshan Yun to help businesses deploy AI into core business.
The enterprise side moves are also encrypted. Tencent stated on 29 October that its AI agent development platform, ADP, allowed businesses to customize, manage, coordinate and use AI agent for customer service, marketing and operations; and Alibaba Cloud had launched Agent Loop and Agent Teams in July to develop, deploy and monitor AI agent in conjunction with the original Agent Run.
It's the same place: the next entrance to AI. The chat is the entrance, the office is the entrance, and it is more likely to be certified as an entrance, the AI agent, who did it without a word.
DeepSeek entered in a different position than the previous ones. It opens the bottom Harness source, models, tools, skills, sessions, sandboxes, all plugins, MIT protocols; the Super-agent prototype is also open source and enters from its own R & D scene, the first of which is Coding Agent, based in the code warehouse.
This position has a direct impact on the application layer. In the past, the moat at the application level was “a better understanding of how to use the model”, and once the bottom model company opened up the Agent framework and best practices, the low capacity of the layer became rapidly thinner. DeepSeek is a new battleground for himself: it used to be with Open AI, Anthropic models, and now with byte, Tencent, Aliby products, contrast scenes, more than anyone who could actually put AI agent into real workflow.
One of the words in this post is DeepSeek's position: “The system has taken over cumbersome session management at the bottom, synchronized with context, allowing humankind to move further from specific details of implementation to higher levels of decision-making and acceptance check.”
This is a definition of a company ' s own product, not an industry judgement. To test it, the future depends on two things: how long the internal example ran, what it ran, and whether the public prototype of publish was able to expand from a GitHub warehouse to more scenes.
Finally, the article did not disclose the scale of the operation of the internal example and did not give the task success rate. The public part is a research prototype, a GitHub warehouse and a set of experimental data based on synthetic loads. Therefore, everything has yet to be tested.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.