A prototype, and labelled as one
DeepSeek published dsh-superagent-prototype on GitHub on 31 October under the MIT licence. By the afternoon of 1 November it had about 6,800 stars and 676 forks.
The repository opens by telling readers what it is not. “Research prototype for reference only, not an official DeepSeek product,” the README says, adding that the code grew out of the team’s day-to-day development work “with much of the engineering done by coding agents”. It warns that the code may still contain bugs and that its style and engineering practices may be inconsistent, and that the published version is not identical to the internal one, nor to the implementation described in an accompanying article, Exploring Super-agent Design: Towards a More Autonomous Agent System. Some designs discussed in that article, it says, are directions for future development and are not implemented here.
What is implemented is narrow. dsh-superagent is a proactive coding agent scoped to one GitHub repository, packaged as a bundle for DeepSeek Harness, the company’s much larger agent framework. It listens to GitHub and, optionally, to Feishu. When someone allowed asks it for work, it opens a work item, works on it in its own agent session and directory, and answers in the original conversation — with a pull request, for example. A web panel shows what needs a person.
The constraint is attention, not capability
The first numbered document in the repository is not about architecture. It is about what the agent costs the people around it.
“Take as little human attention as possible, and give as much value as possible,” the philosophy document states, and then applies that as a test. “Every part of the project is judged by the attention it costs. A message, a notification, a question, a setting, a panel element and a paragraph of documentation each cost someone’s attention, and each has to give more value than that, or it is removed.”
The rules that follow are unusually restrictive for a system marketed on autonomy. The agent acts only when asked directly by a person the team has allowed, or when a written rule permits it — and work nobody asked for “starts in a mode that only records what the agent would have done”. It speaks only when a person must act or can use a result, confirms a request once, then stays quiet while it works. It asks one concrete question at a time. It stops by impact rather than effort: a local change goes as far as a reviewable result, while a change needing a design decision stops with a question. And it reports honestly — a draft is marked unfinished, a failure is reported as a failure, and a limit is never bypassed to avoid asking.
A warning about volume
The same document puts the responsibility on the operator rather than the model, in plain terms. The agent comments, opens pull requests and messages people under accounts a team sees every day, it notes, and “every message it sends costs your colleagues’ attention, and you are responsible for it”.
Its advice is to start narrow — a short allowed list, proactive work recorded only — and widen once what the agent would have done looks right. When the agent is noisy, the document says to make its prompts shorter and stricter or pause it, rather than adding more messages to explain the noise. One instruction is blunter than the rest: “Do not use it to produce volume.” Mass pull requests, broadcast messages and busywork, it says, cost the team more than they give.
What it says about the design argument
DeepSeek is not claiming a capability jump here, and the repository is explicit that the published code trails both its internal version and the paper. What it does publish is a position: that the limit on a repository-wide agent is not what it can do but what it can be allowed to interrupt.
Setup is itself delegated — the README tells users to point their own coding agent at SETUP.md and have it install and configure the thing, asking one short question at a time. The documentation is numbered 000 to 008 and meant to be read in order.
The accompanying article is where the unimplemented ideas live. Until DeepSeek ships any of them into DeepSeek Harness itself — the main repository carries more than 268,000 stars — this remains a design argument with a working demonstration attached, not a product.
