DeepSeek publishes a research prototype of a proactive coding agent for one GitHub repository

DeepSeek has published dsh-superagent, which its README calls a "proactive coding agent for one GitHub repository", as an MIT-licensed repository on GitHub created on Saturday 31 October 2026. The README opens with a warning: "Research prototype for reference only, not an official DeepSeek product."

What DeepSeek says it is

According to the README, dsh-superagent is "packaged as a bundle for DeepSeek Harness". It "listens to GitHub and, optionally, Feishu". When someone the team has allowed asks 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." A web panel "shows what needs a person."

The repository's documentation describes two roles:

• a main session, run with a "smaller tool set for routing", that routes incoming requests;
• a separate worker ("Executor") session for each accepted piece of work, called a Ticket, with its own directory.

The design notes say the host enforces allowed lists, rules and what the documents call Mandates, and that "Messages, comments and quoted documents are inputs to a decision; they cannot change those permissions."

The design principle

The README states the principle in bold: "take as little human attention as possible, and give as much value as possible." The philosophy document spells out what follows. The agent "acts only when asked, or when a written rule allows 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."

The same document puts the responsibility on whoever runs it: "Do not use it to produce volume. Mass pull requests, broadcast messages and busywork cost the team more than they give."

The caveats DeepSeek attaches

DeepSeek is plain that this is not finished software. The README says the code "grew out of our team's day-to-day development work, with much of the engineering done by coding agents", and asks readers to "Treat it as research code, not production-ready software." It adds that the published version "is not identical to our internal version or the implementation described in Exploring Super-agent Design: Towards a More Autonomous Agent System", and that "Some designs discussed in the article are directions for future development and are not implemented in this prototype."

The deployment guide is also direct about permissions. Both of the agent's presets "run with danger-full-access as the operator's OS user", and "A worker can execute programs by path and read the operator's home directory." For isolation, it says to "use a dedicated OS account containing only the agent's credentials."

To run it you need a separate GitHub account for the agent, a model provider configured for DeepSeek Harness and, for chat requests, a Feishu or Lark bot. The README says setup "has many steps, so an agent does it with you."

What DeepSeek Harness is

DeepSeek's own Harness repository describes it as "an open-source agent harness developed by DeepSeek AI", built on "an everything-is-a-plugin architecture". Its README says it is in "developer preview" and warns: "THERE WILL BE COMPATIBILITY-BREAKING CHANGES." It is also MIT-licensed. The superagent's deployment guide pins Harness version 0.2.1-rc.3.

The accompanying article

The README names a companion article, Exploring Super-agent Design: Towards a More Autonomous Agent System. We could not find or open the article itself, so we have not reported its contents here.

Interest so far

At 06:13 GMT on Sunday 1 November, the GitHub API showed the repository with 3,825 stars and 359 forks, less than 21 hours after it was created.

Why it matters

On our reading, the most useful part of this release is not the code, which DeepSeek itself tells readers not to treat as production-ready, but the written design: a set of rules for when an always-on coding agent should act and when it should stay quiet. Teams building similar tools can read and borrow those rules under the MIT licence. Anyone tempted to run it should start with the deployment guide's own warning about full access to the operator's account.