DeepSeek Decided the Bottleneck Was Your Attention, Not the Model

On October 31, the DeepSeek Harness team published a technical note titled Super-agent: Design Exploration Toward a More Autonomous Agent System. A prototype shipped alongside it.
The argument starts somewhere unusual for a model company. It does not begin with capability. It begins with a limit on the human side: a person can follow three to five active sessions before quality collapses. When you run several tasks in parallel, you end up manually switching context and remembering which conversation holds which decision. Single-task ceiling comes from the model. Multi-task ceiling comes from you.
That is a systems problem, and DeepSeek treated it like one.
Router and Worker
The architecture is two layers. A Router runs continuously as a global coordinator. It receives messages from whatever channels you already use, including group chats, GitHub and email, and dispatches tasks. Workers each hold their own context and execute in parallel: reproducing a bug, reviewing a pull request, building a feature. A Worker can call a Subagent when a task needs to fan out further.
Around that sit four mechanics: event-driven dispatch with two-level filtering, asynchronous multi-channel interaction, separation of storage and compute, and an external watchdog.

The event-driven part is where the autonomy claim lives. In the internal instance DeepSeek runs, a team member reporting a software problem in a group chat can trigger the agent without mentioning it. The system recognises a potential defect, reproduces it in the background, and looks for a fix. No one assigned that task. Nothing waited for a prompt.
DeepSeek lists four properties it is aiming for: continuous, meaning one agent across IM and GitHub instead of isolated sessions; efficient, meaning a new instruction does not wait for the previous task to finish; proactive, meaning it watches environment events and acts when something deserves attention while staying silent otherwise; and persistent, meaning always on, plugged into real work environments, and able to recover from failures during long runs.
The example the team gives is instructive. Say you propose plan A in a chat, then change your mind in a pull request comment and propose plan B. A session-bound agent would need to be told. The Super-agent is supposed to notice and adjust direction.
The honest disclaimer
DeepSeek states plainly that this is an engineering exploration, not an announcement of product features. The team also says the design introduces no new theory. It applies classic systems ideas, the kind that have been in operating systems and distributed systems for decades, to agent construction.
That modesty is accurate and also the most interesting thing in the document. The hard part of agents at scale is context management, scheduling, failure recovery and observability, which is to say it is infrastructure work rather than intelligence work.
The numbers that arrived five days later
On November 5, the same team posted community figures for DeepSeek Harness. Measured by users of the official API, daily active users reached 1.54 million on November 4, monthly active users reached 5.48 million, and daily token consumption was around 92 trillion.
Do the division. Ninety-two trillion tokens across 1.54 million daily actives is roughly 60 million tokens per active user per day. No chat application produces that profile. Agents do. A single long-horizon coding task can consume more tokens in an hour than a person types into a chat window in a month.
The plugin numbers matter more for where this is going. Third-party plugins have grown to about 35,000, and roughly 60% of active users have made or installed at least one. DeepSeek says it will tighten plugin API stability and compatibility, improve the plugin security model, and build an official marketplace, with a support programme for plugin authors in preparation.
That is a platform move. Once plugins are the distribution channel, the harness competes on ecosystem rather than on model quality alone.
Why this matters outside DeepSeek
Three things transfer to any team running agents in production.
First, the interface is the existing one. The agent shows up in group chat and pull request comments rather than requiring a new destination. Photoshop never asked users to learn a new app; it lived inside the document. Agent platforms are learning the same lesson.
Second, attention is a resource with a hard limit, and it is the one nobody is scaling. You can buy more GPUs. You cannot buy more of a person's ability to hold four threads of context. Every architecture decision that moves coordination from the human to the system buys back the scarcest input in the loop.
Third, self-hosting the coordination layer is where the operational risk lands. An agent that acts without being mentioned needs a watchdog, because the failure mode is silent: it does something reasonable that nobody asked for, at 3 a.m., in a repository that matters.
The part still missing
The document does not address cost control. A router that dispatches work and workers that reproduce bugs on their own schedule will spend money on its own initiative. The 92 trillion token figure suggests that spending is already large; how much of it produced work anyone wanted is not disclosed.
Nor does it address the review problem. If an agent can open a pull request in response to a chat message, someone has to review it. DeepSeek's framing is that humans move up to goal-setting and acceptance. That works when acceptance is cheap to judge. In most engineering organisations, judging whether a fix is correct is the expensive part, and it does not shrink because the agent got faster.
The part that sounds like an operating system
Strip the AI vocabulary and the Super-agent document reads like a paper on job scheduling. A long-running coordinator receives events, classifies them, and dispatches work to isolated contexts. Storage and compute scale separately so a stalled task does not hold memory a live one needs. An external watchdog restarts anything that stops reporting. Two-level filtering exists so the system does not wake a human for every message in a group chat.
Those are solved problems in distributed systems. They are unsolved in agent products, because most agent products were built as a chat session with tools attached, and a chat session has no concept of concurrent work.
The interesting consequence is organisational. If the coordination layer genuinely absorbs session management, the job of the person in the loop changes from operating a tool to writing acceptance criteria. DeepSeek frames the human as moving up to goal-setting and verification. In practice that means the skill that matters shifts from prompt writing to knowing what done looks like, which is a management skill and considerably harder to teach.
DeepSeek's own summary is that this way of working could become normal for human-machine collaboration. The mechanism is credible. The open question is whether organisations will reorganise around a coordinator that never sleeps, or simply bolt one onto processes designed for people who do.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.