文 | 影子备忘录
10月31日晚，DeepSeek Harness团队在官方公众号发了一篇技术文章。同一天，他们的内部Super-agent实例正在做一件事：没有人@它，没有人新建会话，也没有人把某条群聊消息“搬运”到GitHub。
它自己在群聊里识别到一个软件问题的线索，在独立目录里复现了问题、生成了修复方案，然后回到群聊交付了一条可评审的Pull Request。
而今天，我更想聊的是它背后一些隐形判断：AI单任务能力已经跨过了“够用”的线，真正的瓶颈不是模型不够强，是人的注意力不够分。
与此同时，另一条线在资本市场上同步推进：据公开报道和市场消息，DeepSeek完成了约1000亿元人民币的融资，腾讯和宁德时代是出资最多的投资方，投前估值维持在约5000亿元，中信证券正在协助筹备科创板IPO，计划最快2026年末递表、2027年挂牌。
不过DeepSeek及腾讯、宁德时代等相关方未公开回应或官宣，最终金额和交易细节仍存在变动可能。
技术叙事和资本叙事，在这一周交汇。
Super-agent承认“人管不过来”
大多数Agent产品的叙事是“帮你做更多”。写代码、查资料、发邮件、做表格，功能列表越拉越长。
但有一个数字，所有做Agent的人心里都清楚，只是很少被摆到台面上：一个人通常只能同时跟进3到5个活跃会话。
这意味着过去一年Agent产品的繁荣，很大程度上是在给人“加会话”。
每多一个Agent，人就多一个需要创建、投喂、盯进度、搬结果的窗口。
Agent越多，人越像一个低带宽的消息总线，在不同会话之间反复切换、遗忘、重新加载上下文。
模型能力在涨，人的管理带宽没涨。
Super-agent的Router/Worker双层架构，本质上是在承认这件事，然后把它从人身上拿走。
Router常驻在线，统一接住群聊、GitHub、邮件进来的消息，维护全局上下文；Worker各自隔离上下文并行干活，结果由Router统一交付。
用户的感觉是同一个Agent，背后则是一组Agent在同时工作。
更值得琢磨的是团队的设计哲学。他们明确说这套架构没有提出新理论，而是把计算机系统里成熟的调度、缓存、存算分离、看门狗思路搬进了Agent系统。
它不是在让模型变聪明，而是在让Agent系统变得“可运营”。
一个只能单次演示的系统，永远需要人站在旁边当调度员；一个可运营的系统，才敢把越来越多的任务放进去持续跑。
用大白话翻译：过去是人管Agent，现在系统开始管Agent，人则往后退了一步。
从“比模型”到“比谁能把Agent管住”
如果你只盯着DeepSeek这一篇文章，可能会觉得这是一个孤立的技术探索。但把时间线拉长一点，会发现这不是一家公司在转向。
2026年上半年的一份行业复盘写得非常直白：模型竞争的重点“不再只是参数规模和榜单排名，而是更综合的系统能力”，即长上下文、工具调用、成本控制、Agent任务执行能力等，这些正在取代“谁家参数更大”成为新的计分板。
另一份行业观察则直接说，“训练参数量谁更大”的旧胜负手正在失效，“VLA + 世界模型 + Agent端到端工程闭环”成了新的胜负手。
大洋彼岸的动作更早、更系统。OpenAI把Sessions、Agents SDK、Tracing、Evals做成一等开发面；Anthropic把Memory、Compaction、Context Editing、Managed Agents逐步补齐；Google把ADK、Agent Runtime、Memory Bank、Agent Gateway组织成完整平台。
方向都指向同一个目标：把Agent从一个“模型调用”变成一套“可运行的系统”。
为什么这个转向是必然的？因为Agent一旦从单轮问答走向长链条任务执行，Token消耗会随任务拆解、工具调用和上下文积累快速放大。
Agent带来上下文重复调用，大量Token被浪费；产业链利润向上游集中，下游还没找到稳定现金流场景却已经先行投入。
当Token单价降了99%，但Agent把消耗量放大了几十倍，账单反而更贵了。
“让模型多干活”不再免费，系统效率就变成了真金白银的竞争维度。
所以Super-agent那组实验数据或许值得认真看：约300条消息、20项工作请求的测试中，任务完成时间p50为12.4分钟、p95为15.7分钟，带Subagent工具的单会话方案在中性提示下p95是29.5分钟。
去掉订阅直通后，Router决策次数增加约30%，读缓存Token从2323万涨到3177万。
它拿更多Token换时间和稳定性。这恰恰说明，系统层竞争的本质不是“省Token”，它更像是“用工程手段让每一分算力花得更值”。
Manus和扣子空间的对比也能佐证同一件事。Manus复杂任务成功率92%，扣子空间81%，差距在多步纠错和长程记忆维持的场景里被放大。
但两边都撞上了同一堵墙：上下文一长就“失忆”，多工具串联就卡住，中间步骤不可回溯。
这不是模型不够聪明，是系统没有把上下文当成一种需要被管理的基础设施来对待。
千亿融资买的是“系统化”的入场券
再把镜头拉到资本侧。
这笔融资的细节，比“纪录”两个字更有信息量。募资目标最初500亿，经多轮上调到约1000亿；投前估值锚在5000亿；腾讯两轮各出100亿，宁德时代体系本轮出资100亿，较首轮的50亿翻倍；中信证券进场筹备科创板，高瓴创投原合伙人严文韬已出任CFO。
资金投向写得很清楚：模型训练、算力建设、人才招募，支撑全球竞争。
但真正值得拆的不是金额，是出资方的结构在说话。
腾讯连续两轮重仓，说明大厂参与AI竞争的方式正在从“只自研”变成“自研+入股头部系统”。
一边混元在追，一边用资本锁定DeepSeek的基础模型和开源全栈能力，既补短板，也提前买这张牌桌的入场券。
宁德时代的逻辑更产业侧：当AI系统开始接管真实工作流，算力、能源和数据中心的约束会从成本项变成竞争壁垒，产业资本比纯财务资本更早闻到这一层。
更值得注意的信号来自运营商的入局。据The Information报道，中国移动和中国电信出现在本轮近30家投资方名单中，单家金额达数十亿元级别。
运营商手里握着的恰好是模型公司最难靠自己补齐的东西：全国性IDC与电力、骨干网带宽、智算集群、覆盖全国的政企渠道。
DeepSeek近期开源昇腾适配组件、推进国产算力路线，与运营商“国芯训国模”的方向形成合拍。这不再是“财务投资”，是算力基础设施和模型生态在资本层面做绑定。
所以这千亿融资，更像是国产AI头部竞争的门槛从“能不能训出好模型”抬高到了“能不能同时搞定算力、系统、产业渠道和资本市场”。
上海证券报引述的分析说得很直白：资金、算力、深度智能，以及“模A一体”布局的市场进展，是拉开差距的四个关键变量；这轮融资不是终点，通过上市继续融资是必然要走的路。
IPO铺路是这条逻辑的延长线。5000亿投前估值对应的是二级市场对“系统化能力”的定价预期，而不是对单一模型参数的定价。
资本都在为一件事投票：下一阶段的赢家，是能把模型、Agent系统、算力基础设施和组织执行方式整合成一个可持续运行整体的公司。
“验收”比“执行”更难
回到那句最容易被当作口号滑过去的话：人从执行细节上移到决策与验收。
这句话的重量，不在“人轻松了”，而在“人被要求做一件更难的事”。
过去，执行和监督是同一件事的两面。你写代码的时候知道每一行为什么这么写，你盯进度的时候知道卡在哪一步。
现在系统把执行折叠进内部，人看到的是一份“已经整理好的结果”。验收的前提，是你有能力判断结果对不对，可是当过程不再经过你的手，这种判断力从哪来？
这就演变成了一个组织问题。Super-agent的哲学文档里写了“尽可能少占用人注意力”，也设计了“未完成成果明确标注”“需要决策时结构化提问”这些机制。
但“减少信息过载”和“保留可审计性”之间有一条很窄的缝。一个系统在20项请求上跑通，和在2000项并发请求上不漂移、不失控、每一份结果都能追溯到依据，是两种难度。
更底层的张力在于：当Agent系统开始自己拆任务、自己派活、自己推进，人的“决策”就不再是“告诉AI做什么”，而是“判断AI选的方向对不对、交付的东西能不能用”。
前者是执行层的技能，后者是判断层的能力。提示词可以学，验收力需要领域判断和责任意识，学起来慢得多。
所以目前行业一些人不断强调“AI在解放人类”的说法，我个人是不太认同的。更准确的说法是：执行层的带宽正在被系统接管，判断层的责任正在被重新压回人身上。
这种转变，恰恰是把更难的事儿，留给了人。
边界说清楚，方向才看得清
技术侧，Super-agent是一套面向工程实践的架构探索，配套原型以MIT协议开源，但团队自己强调“并非正式产品”，实验基于合成负载与简化环境，是初步参考。
方向有工程支撑，离大规模长周期混跑还有距离。
资本侧，约1000亿融资、腾讯宁德时代为最大出资方、2027年初IPO的时间表，均来自公开报道和市场消息，相关方未公开回应或官宣。
趋势判断可以成立，事实表述需要留出精度。
但方向本身已经足够清楚：AI竞争的下一章，比的不是谁家模型在榜上多一分，而是谁家的系统能让Agent稳定地、持续地、在真实工作流里把事做完。
千亿融资和IPO铺路，买的是这张系统化竞赛的入场券；Super-agent的Router/Worker架构，是这张入场券上印着的第一个工程注脚。
而人需要回答的问题，不在架构图里：当执行不再经过你的手时，你还能不能说出“这个结果是对的”。
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。
A shadow memo.
On the evening of 31 October, the DeepSeek Harness team published a technical article in the official WeChat Official Account. On the same day, their internal Super-agent example was doing one thing: no one @ it, no one created a new session, and no one was taking a group of people to GitHub to talk about "porting."
It itself identified a clue to a software problem in the group chat, recreated the problem in the independent catalogue, generated a repair programme and then returned to the group to deliver a retrievable Pull Request.
And today I would like to talk more about some invisible judgments behind it: AI's single-manufacturing capability has crossed the “sufficient” line, and the real bottleneck is not that the model is not strong enough, but that human attention is not enough.
At the same time, another line is advancing simultaneously in the capital market: According to public and market reports, DeepSeek has completed financing of about RMB 100 billion, Tencent and Ningdee were the most funded investors and the pre-investment valuation was maintained at around RMB 500 billion, and Chinese securities are assisting in the preparation of the IPO, which is planned to be delivered by the end of 2026 and put up in 2027.
However, the parties concerned, such as DeepSeek and Tencent and the Ningde era, have not responded publicly or have declared that the final amount and details of the transaction are likely to change.
The technology narrative and the capital narrative converged this week.
Super-agent admits that "there's no human control."
Most of the narratives of Agent's products are “do more for you”. Write codes, check information, e-mail, make tables, and the list of functions is drawn up longer.
But there is one figure that all those who do Agent know well, but are rarely put on the table: one can usually follow only three to five active sessions at the same time.
This means that the boom in Agent's products over the past year has been largely “talking”.
For each additional Agent, there is an additional window that needs to be created, fed, followed by progress and results.
The more Agent, the more people look like a low-bandwidth news bus, the more they switch, forget, and reload context.
Modelling capacity is growing and human management bandwidth is not increasing.
The Router/Worker two-tiered structure of Super-agent was essentially to acknowledge the matter and then take it from people.
Router is based on the online network, which integrates the news of the group chat, GitHub, e-mails, and maintains global context; Work is working in isolation from context in parallel, with the results delivered by Router.
The user senses the same Agent, behind which a group of Agent works at the same time.
More interesting is the philosophy of team design. They made it clear that the architecture did not present a new theory, but instead moved mature movement plans, cache, memory separation, and door dog thinking into the Agent system.
It's not making the model smart, it's making the Agent system operational.
A single-demonstration system that always requires people to stand next to it as dispatchers; an operational system that dares to put more and more tasks in it and keep running.
Translating in big white: Agent used to be a man, and now the system is running Agent, and one step back.
From "By Model" to "Better than Who can hold Agent."
If you only look at DeepSeek, you might think it's an isolated technical exploration. Longer time lines, however, will reveal that this is not a turn of company.
An industry revisit in the first half of 2026 was very straightforward: the focus of model competition “is no longer just the size of the parameters and ranking of the list, but rather a more integrated system capacity”, i.e. long context, tool call, cost control, Agent's ability to perform tasks, etc., which is replacing “who is larger” as a new scoreboard.
Another industry observation said that the old winner “training parameter count who is bigger” was failing, and the “VLA + World Model + Agent End-to-End Project Closer” became the new winner.
It's much earlier and more systematic across the ocean. OpenAi has made a first-class development of Sessions, Agents SDK, Tracing, Evals; Anthropic has gradually completed Memoory, Company, Context Editing, Managed Ages; and Google has organized ADK, Agent Runtime, Memoory Bank, Agent Gateway.
The direction is directed towards the same goal: to transform Agent from a "model call " to a " running system".
Why is this shift inevitable? As Agent moves from a single-cycle question and answer process to a long chain mission, token consumption is rapidly amplified by the dismantling of the mission and the accumulation of tool call and context.
Agent brought with him contextduplicate call, and a large number of Tokens were wasted; the chain profits were concentrated upstream, while the stable cash flow scene was not found downstream but had been invested first.
When Token unit prices fell by 99 per cent, but Agent scaled up consumption several dozen times, the bill became more expensive.
“Let the model work more” is no longer free, and the efficiency of the system becomes the competitive dimension of real gold and silver.
So the Super-agent set of experimental data may be worth looking at: about 300 messages, 20 job requests tested, with mission completion time of 12.4 minutes at p. 50, 15.7 minutes at p. 95, and a single session program with a Subagent tool with a neutral hint of 29.5 minutes at p. 95.
After removing direct subscriptions, Router increased the number of decisions by about 30 per cent, reading cacheToken from 2.323 million to 317.7 million.
It takes more Token for time and stability. This shows precisely that the nature of system-level competition is not “the province of Token”, more like “engineering to make every fraction more valuable”.
A comparison of Manus and Coze space can also support the same thing. The Manus complex mission success rate92 per cent and the Coze space 81 per cent have widened the gap in the scenes of multiple error correction and long-range memory maintenance.
However, both sides hit the same wall: context was “mnesia” for a long time, multiple tools were stuck and the intermediate steps were not traceable.
This is not that the model is not smart enough; it is that the system does not treat context as an infrastructure that needs to be managed.
Hundreds of billions of dollars of finance buy "systematized" tickets.
And pull shot to the capital side.
The details of this financing are more informative than the word “record”. The fund-raising target was initially $50 billion, which was revised to about $100 billion through multiple rounds; the pre-investment valuation anchor was 500 billion; the Tencent two rounds were $10 billion each; the Ninde-era system was spending $10 billion each, twice as much as the first round; and the CFO had been created by the Sino-Cultures in preparation for the launch of the Sino-Sino-Sino-Sino-Sino-Sino-Sino-Sino-Sino-Sino-Sino partnership.
The investment is well written: model training, calculus-building, talent recruitment to support global competition.
But what really deserves to be broken is not the amount, but the structure of the donor.
Tencent is two successive rounds of heavy silos, indicating that the way in which large plants compete in AI is moving from “self-study” to “self-study + equity head system”.
The mixer is chasing, while capital is being used to lock down DeepSeek ' s basic model and open-source whole house capacity, both to fill the board and to buy tickets to this table in advance.
The logic of the Ninde era is more industrial: as AI systems begin to take over real workflow, the constraints of computing, energy and data centres can shift from cost items to competitive barriers, and industrial capital can smell this layer earlier than pure financial capital.
More noteworthy signals come from the operator's entry. According to The Information, China's mobile and Chinese telecommunications appear on the list of nearly 30 investors in the current round, amounting to billions of yuan each.
Operators hold exactly what model companies have the hardest to fill on their own: national IDCs and electricity, backbone bandwidth, smart clusters, and government channels covering the country.
DeepSeek's recent open source retrofitting components, boosting the national production calculator route, is co-opted with the operator's “ChipTrain Model”. This is no longer a “financial investment”, it is a capital-level binding of computing infrastructure and model ecology.
So this multibillion-dollar finance, more like the threshold of national production AI head-to-head competition, has been raised from “can we get a good model” to “can we get together with calculus, systems, industrial channels and capital markets”.
The analysis cited in Shanghai Securities is straightforward: finance, calculus, deep intelligence, and market progress in the “mode-One” layout are the four key variables that open the gap; this round of financing is not the end, and continued financing through listing is the inevitable path.
The IPO paved the road as an extension of this logic. The $500 billion pre-position valuation corresponds to the pricing expectations for “systemic capabilities” in secondary markets rather than for single model parameters.
Capital is voting for one thing: the next is the winner who can integrate models, Agent systems, computing infrastructure and organizational execution into a sustainable whole company.
"acceptance check" is more difficult than "implementation"
Returning to that phrase would be the easiest thing to slip over as a slogan: people move from implementation details to decision-making and acceptance check.
The weight of the sentence is not “human ease”, but “human being asked to do something harder”.
In the past, implementation and monitoring were two sides of the same coin. You know what every behavior says when you write codes, and you know where the cards are when you look at progress.
Now the system folds implementation into the interior, and what one sees is an “outcome that has been sorted out”. The premise of acceptance check is that you are capable of judging the results, right?
This has evolved into an organizational issue. Super-agent's philosophy document contained “as little attention as possible” and also designed the mechanisms “unfinished results with clear indications” “structured questions when decisions are needed”.
There is, however, a narrow gap between “reduce information overload” and “retain auditability”. It is difficult for a system to run through 20 requests, and for 2,000 simultaneous requests not to drift, not to lose control, and to trace each result back to its basis.
The bottom line of tension is that when the Agent system starts dismantling its own tasks, sending its own activities, and pushing itself forward, human “decision-making” ceases to be “to tell AI what to do” but “to judge whether or not it chooses the right direction, whether or not the delivered thing works”.
The former are skills at the executive level, while the latter are competencies at the judgement level. prompt can learn, and acceptance check needs a sense of area judgement and responsibility and learns much slower.
So now some people in the industry are constantly emphasizing that “AI is liberating humanity”, and I personally do not agree. More precisely, the bandwidth of the executive layer is being taken over by the system and the responsibility of the adjudicative layer is being re-pressed.
It is precisely this shift that leaves more difficult things behind.
The boundaries are clear, the direction is clear.
On the technical side, Super-agent is a set of structural explorations oriented towards engineering practice, with prototypes with open-source MIT protocols, but the team itself emphasizes “non-formal products”, with experiments based on synthetic loads and simplified environments as a preliminary reference.
The direction is engineering-supported and is still a distance from large-scale long cycles.
The capital side, about $100 billion in financing, the TencentNingd era as the largest contributor, and the IPO schedule at the beginning of 2027, all came from public and market information, with no public response or public announcement by the parties concerned.
Trends can be judged and the presentation of facts needs to be refined.
But the direction itself is clear enough: the next chapter of AI’s competition is one point more than the model of no one, and whose system allows Agent to do it in a stable, continuous and real workflow.
Hundreds of billions of dollars of financing and IPO paved the way for a ticket to this systematic competition; the Router/Worker structure of Super-agent was the first project footnote stamped on this ticket.
The question that one needs to answer is not in the chart: can you say, “This is the right result” when the execution no longer passes through your hand?
Auto-fetched and translated by AI Atlas. All rights belong to the original author.