DeepSeek Harness team releases Super-agent architecture

By | Shadow Memo

On the evening of October 31, the DeepSeek Harness team published a technical article on its official WeChat account. On the same day, their internal Super-agent instance was doing something: no one @-mentioned it, no one started a new session, and no one "carried" a group chat message over to GitHub.

On its own, it identified a clue about a software problem in the group chat, reproduced the problem in an independent directory, generated a fix, and then returned to the group chat to deliver a reviewable Pull Request.

Meanwhile, another thread was advancing in parallel in the capital markets: according to public reports and market chatter, DeepSeek completed about 100 billion yuan in financing, with Tencent and CATL as the largest investors, a pre-money valuation held at about 500 billion yuan, and CITIC Securities assisting in preparing a STAR Market IPO, with plans to file as early as late 2026 and list in 2027.

However, DeepSeek and related parties such as Tencent and CATL have not publicly responded or made an official announcement, and the final amount and transaction details may still change.

The technology narrative and the capital narrative converged this week.

Super-agent admits "people can't keep up"

The narrative of most Agent products is "help you do more." Writing code, looking up information, sending emails, making spreadsheets—the feature list keeps getting longer.

But there is one number that everyone building Agents knows in their heart, yet it is rarely put on the table: a person can usually only keep up with 3 to 5 active sessions at the same time.

This means that the prosperity of Agent products over the past year has largely been about "adding sessions" for people.

For every additional Agent, a person gains another window that needs to be created, fed, monitored for progress, and have results moved out of.

The more Agents there are, the more a person resembles a low-bandwidth message bus, repeatedly switching between different sessions, forgetting, and reloading context.

Model capabilities are rising, but human management bandwidth is not.

Super-agent's Router/Worker two-tier architecture is essentially an admission of this fact, and then taking it away from people.

The Router stays online permanently, uniformly catching messages coming in from group chats, GitHub, and email, and maintaining global context; Workers each isolate their context and work in parallel, with results delivered uniformly by the Router.

To the user, it feels like the same Agent, but behind the scenes a group of Agents is working simultaneously.

What is more worth pondering is the team's design philosophy. They explicitly said that this architecture does not propose a new theory, but rather moves mature ideas from computer systems—scheduling, caching, storage-compute separation, watchdog—into the Agent system.

It is not making the model smarter; it is making the Agent system "operable."

A system that can only be demonstrated once will always need a person standing beside it as a dispatcher; only an operable system dares to put more and more tasks into it and keep them running.

In plain terms: in the past, people managed Agents; now the system has begun to manage Agents, and people have taken a step back.

From "comparing models" to "comparing who can keep Agents under control"

If you only look at this one DeepSeek article, you might think it is an isolated technical exploration. But stretch the timeline a bit longer, and you will find that this is not one company changing direction.

An industry review written in the first half of 2026 put it very bluntly: the focus of model competition is "no longer just parameter scale and leaderboard rankings, but more comprehensive system capabilities," namely long context, tool calling, cost control, Agent task execution capability, and so on, which are replacing "whose parameters are bigger" as the new scoreboard.

Another industry observation said directly that the old deciding factor of "who trains more parameters" is failing, and "VLA + world models + Agent end-to-end engineering closed loop" has become the new deciding factor.

On the other side of the ocean, the moves came earlier and more systematically. OpenAI made Sessions, Agents SDK, Tracing, and Evals first-class development surfaces; Anthropic gradually filled in Memory, Compaction, Context Editing, and Managed Agents; Google organized ADK, Agent Runtime, Memory Bank, and Agent Gateway into a complete platform.

All directions point to the same goal: turning an Agent from a "model call" into a "runnable system."

Why is this shift inevitable? Because once an Agent moves from single-turn Q&A to long-chain task execution, Token consumption rapidly expands with task decomposition, tool calls, and context accumulation.

Agents bring repeated context calls, and large amounts of Tokens are wasted; profits in the industry chain concentrate upstream, while downstream has already invested ahead of time without yet finding stable cash-flow scenarios.

When the unit price of Tokens has dropped 99%, but Agents have multiplied consumption dozens of times over, the bill actually gets more expensive.

"Making the model do more work" is no longer free, and system efficiency has become a competition dimension measured in real money.

So that set of experimental data from Super-agent may be worth a careful look: in a test of about 300 messages and 20 work requests, task completion time was p50 12.4 minutes and p95 15.7 minutes, while the single-session solution with the Subagent tool under neutral prompts had a p95 of 29.5 minutes.

After removing the subscription pass-through, Router decisions increased by about 30%, and read-cache Tokens rose from 23.23 million to 31.77 million.

It trades more Tokens for time and stability. This precisely shows that the essence of competition at the system layer is not "saving Tokens"; it is more like "using engineering means to make every unit of compute deliver more value."

The comparison between Manus and Coze Space also corroborates the same point. Manus has a 92% success rate on complex tasks, while Coze Space has 81%, and the gap widens in scenarios involving multi-step error correction and long-horizon memory maintenance.

But both sides hit the same wall: once the context gets long, they "forget"; when multiple tools are chained, they get stuck; and intermediate steps cannot be traced back.

This is not because the model is not smart enough; it is because the system does not treat context as a kind of infrastructure that needs to be managed.

The hundred-billion financing buys an entry ticket to "systematization"

Now let's turn the lens to the capital side.

The details of this financing round carry more information than the word "record." The fundraising target was initially 50 billion, and after multiple upward revisions it reached about 100 billion; the pre-money valuation was anchored at 500 billion; Tencent contributed 10 billion in each of two rounds, and the CATL system contributed 10 billion this round, double its 5 billion in the first round; CITIC Securities came in to prepare for the STAR Market, and Yan Wentao, former partner at Hillhouse Venture Capital, has taken the role of CFO.

The use of funds is stated clearly: model training, compute infrastructure, and talent recruitment, to support global competition.

But what is truly worth unpacking is not the amount, but what the structure of the investors is saying.

On one side Hunyuan is catching up, while on the other side capital is being used to lock in DeepSeek's foundation models and open-source full-stack capabilities, both shoring up weaknesses and buying an entry ticket to this table in advance.

A more noteworthy signal comes from the entry of telecom operators. According to The Information, China Mobile and China Telecom appear on the list of nearly 30 investors in this round, each with an amount in the billions of yuan.

What the operators hold is exactly what model companies find hardest to fill in on their own: nationwide IDCs and power, backbone network bandwidth, intelligent computing clusters, and nationwide government and enterprise channels.

English text is being prepared.

An analysis quoted by the Shanghai Securities News puts it bluntly: capital, compute, deep intelligence, and market progress in the "model-A integration" layout are the four key variables that create the gap; this financing round is not the endpoint, and continuing to raise funds through a listing is an inevitable path.

Paving the way for an IPO is the extension of this logic. The 500 billion pre-money valuation corresponds to the secondary market's pricing expectations for "systematization capability," not for the parameters of a single model.

Capital is all voting for one thing: the winner in the next stage will be the company that can integrate models, Agent systems, compute infrastructure, and organizational execution into a whole that runs sustainably.

"Acceptance" is harder than "execution"

Back to the sentence that is easiest to let slide by as a slogan: people move up from execution details to decision-making and acceptance.

The weight of this sentence lies not in "people have it easier," but in "people are being asked to do something harder."

In the past, execution and supervision were two sides of the same thing. When you wrote code, you knew why each line was written that way; when you watched progress, you knew which step was stuck.

Now the system folds execution inside, and what people see is "an already organized result." The premise of acceptance is that you are able to judge whether the result is right, but when the process no longer passes through your hands, where does that judgment come from?

This then becomes an organizational problem. The Super-agent philosophy document writes about "occupying as little of people's attention as possible," and also designs mechanisms such as "clearly marking unfinished results" and "structured questioning when a decision is needed."

But there is a very narrow gap between "reducing information overload" and "preserving auditability." A system running through 20 requests, and a system that under 2,000 concurrent requests does not drift, does not spin out of control, and can trace every result back to its basis, are two different levels of difficulty.

The deeper tension is this: when the Agent system starts breaking down tasks itself, assigning work itself, and pushing things forward itself, human "decision-making" is no longer "telling the AI what to do," but "judging whether the direction the AI chose is right and whether what is delivered is usable."

The former is a skill at the execution layer; the latter is a capability at the judgment layer. Prompts can be learned; acceptance capability requires domain judgment and a sense of responsibility, and it is much slower to learn.

This shift is precisely leaving the harder things to people.

Only when the boundaries are clear can the direction be seen clearly

On the technical side, Super-agent is an architectural exploration oriented toward engineering practice, and the accompanying prototype is open-sourced under the MIT license, but the team itself stresses that it is "not a formal product," and that the experiments are based on synthetic loads and a simplified environment, serving as a preliminary reference.

The direction has engineering support, but there is still a distance from large-scale, long-cycle mixed running.

On the capital side, the roughly 100 billion financing, Tencent and CATL as the largest investors, and the timetable for an IPO in early 2027 all come from public reports and market news, and the relevant parties have not publicly responded or made an official announcement.

Trend judgments can hold, but factual statements need to leave room for precision.

But the direction itself is already clear enough: in the next chapter of AI competition, what is compared is not whose model scores one point higher on the leaderboard, but whose system can make Agents reliably, continuously, and in real workflows get things done.

The hundred-billion financing and IPO groundwork buy an entry ticket to this systematization race; Super-agent's Router/Worker architecture is the first engineering footnote printed on that ticket.

And the question people need to answer is not in the architecture diagram: when execution no longer passes through your hands, can you still say "this result is correct"?