DeepSeek Superagent Prototype Is Open Source

DeepSeek’s Harness team has published a DeepSeek superagent prototype on GitHub: a coding agent that sits on one repository, takes work from approved people and answers with pull requests. The code went up on 31 October under the MIT licence, and the repository’s README opens with a warning in bold: “Research prototype for reference only, not an official DeepSeek product.”

The release comes with a long design essay from the same team, published the same day on its official WeChat account. Its argument is that the limit on coding agents is shifting from what the model can do to how much a person can keep track of. The team says that, by its own observation, one person can usually follow only three to five active agent sessions at once. The prototype is its attempt to take that juggling away from people.

In brief

DeepSeek’s Harness team has open-sourced dsh-superagent, a research coding agent for one GitHub repository, with optional Feishu chats.
It works only for approved people, handles each job in its own session and replies in the original thread, for example with a pull request.
It is not an official DeepSeek product, may contain bugs and differs from the team’s internal version.

How the DeepSeek superagent works

In plain terms, dsh-superagent is “a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness”, the company’s agent app, which DeepSeek says is open source and in public preview worldwide. Once set up, it listens to GitHub and, optionally, to Feishu chats.

“When someone allowed asks the agent for work, it opens a work item, works on it in its own agent session and directory, and answers in the original conversation, for example with a pull request,” the README says. A web panel, opened as “Superagent” in the DeepSeek Harness sidebar, lists first whatever needs a person.

The essay describes the design behind it as two layers. A single always-on “Router” takes in messages, keeps the overall context and hands out jobs. Separate “Workers” do the jobs in parallel, each in its own isolated session, and report back. The team says its internal version, built on DeepSeek Harness, already runs continuously. It can carry one task from a group chat to a GitHub pull request review. It also watches feedback in a public group chat and, without anyone mentioning it, finds the cause of a defect and submits a fix as a pull request.

Built to stay quiet unless a person is needed

The DeepSeek superagent’s guiding line is to “take as little human attention as possible, and give as much value as possible.” In practice, the README says, the agent “acts only when asked or when a rule the team wrote allows it, and speaks only when someone must act or can use a result.”

The setup guide is meant to be read by a coding agent, which then walks the operator through the install one question at a time. Its defaults are cautious:
– The agent gets its own GitHub account, “never the operator’s own”.
– At first only the operator may ask it for work.
– Unprompted work starts in an “observe” mode that records what the agent would have done without doing it.
– Auto-merge is off: “people merge until the agent’s pull requests have been right for a while.”

The project’s philosophy note is blunt about misuse: “Mass pull requests, broadcast messages and busywork cost the team more than they give.”

The team’s own tests are small and synthetic

The essay includes early experiments, which the team calls preliminary. Each run fed about 300 chat and CI messages, around 20 of them real work requests, to four set-ups, five times each. Coding itself was replaced by a tool that only used up time, so the tests measured scheduling, not code quality.

On those tests, the DeepSeek superagent’s typical (p50) task took 12.4 minutes and its slow tail (p95) 15.7 minutes. A single agent session given a helper-agent tool and a neutral prompt never delegated in any of five runs, and its p95 reached 29.5 minutes. The superagent used more cached tokens than that single session. Removing a feature that sends follow-up messages straight to the right Worker raised Router decisions by about 30% and cached-read tokens from 23.23 million to 31.77 million, with completion times about the same.

The repository is also clear about its limits. Much of the engineering was “done by coding agents”, the code “may still contain bugs”, and it is “not identical to our internal version”. Some ideas in the essay “are not implemented in this prototype”. The essay itself lists open problems: memory over months, agents talking to other agents, and security. It warns that if a cloud-hosted agent loop were compromised through a server breach or prompt injection, every device connected to it could be at risk.

The design spends tokens to save people’s attention

What stands out is that DeepSeek is publishing a way of working rather than a model. Releases such as DeepSeek-V4.1-Flash-1017 are model updates, judged on scores and on outside safety checks like Fangcun’s evaluation. This release asks to be judged on something harder to measure: whether a team trusts an agent enough to let it work in its threads without someone watching every step.

The trade-off is explicit. The superagent finishes work faster in the team’s tests by running several sessions, which costs more tokens than one session doing everything itself. That is a bet that people’s time, not compute, is the scarce resource. It is the same contest playing out in Factory’s Droid harness, Replit’s cloud agents and Microsoft’s Copilot harness: the harness around the model now matters as much as the model.

The most useful part for other teams may be the restraint. Observe mode, a short allowed list and no auto-merge mean the agent earns more freedom only after its record looks right, and the “would have done” log becomes a test set. The weakness is the one the team names itself. An agent that listens to open chats is exposed to whatever is posted there, and a prototype partly written by coding agents is not yet something to point at a sensitive codebase.

With more than 1,500 GitHub stars at the time of writing, interest is high. The open question is whether DeepSeek turns any of this into a supported DeepSeek Harness feature, or leaves the DeepSeek superagent as reference code that others adapt.