The GitHub organization deepseek-ai has published dsh-superagent-prototype, an MIT-licensed bundle that implements a proactive coding agent for a single repository on DeepSeek Harness. The repository is research code teams can inspect and adapt—not a consumer product launch or a new model release from DeepSeek.
Not an official DeepSeek product
The README opens with a disclaimer the project maintainers ask readers to treat as binding: "Research prototype for reference only, not an official DeepSeek product." The code grew out of day-to-day development work, with much of the engineering done by coding agents. It may contain bugs, and its style and practices may be inconsistent. The published tree is adapted for release; it is not identical to DeepSeek's internal version or to every design in the article Exploring Super-agent Design: Towards a More Autonomous Agent System, and some ideas in that article are future directions not implemented in this prototype.
Key facts
| Item | Detail |
|---|---|
| Publisher | deepseek-ai on GitHub |
| License | MIT |
| Status | Research prototype, not an official product |
| Works with | One GitHub repo, optional Feishu |
| Runs on | DeepSeek Harness |
What DeepSeek published
The project name in the README is dsh-superagent. It is a proactive coding agent scoped to one GitHub repository, packaged as a bundle for DeepSeek Harness. Setup is intentionally multi-step; the maintainers expect a human to work through it with a coding agent by asking it to read SETUP.md in the repository.
How it works on GitHub and Feishu
The agent listens to GitHub and, optionally, Feishu. When someone on an allowed list asks for work, it opens a work item (a Ticket in the domain model), 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 still needs a person. On GitHub, requests arrive when an allowed account mentions the agent's GitHub login in an Issue or comment; the deployment guide says polling runs on a one-minute interval by default.
The attention-saving philosophy
The README states the operating principle plainly: "take as little human attention as possible, and give as much value as possible." Because an agent that can act for a whole team can also waste a whole team's attention, it 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.
Who can ask—and what permissions mean
Policy settings define allowed lists per Source. Under each Source, operators add the GitHub accounts and Feishu users who may ask for work; until a list has someone, no request can become a Ticket. With Feishu enabled, a chat can sync the list at startup and on demand. The deployment documentation also notes that editing an Issue body to add a mention is not received—you must mention the agent in a comment.
What you need to run it
Running the stack requires a dedicated DeepSeek Harness profile (called superagent in the docs), a separate agent GitHub account that comments, pushes branches and opens pull requests through its own GitHub CLI credentials, and optionally a Feishu bot. Prerequisites include Node.js 22.19+ or 24+, pnpm, git 2.32+, Python 3.8+, tmux, and GitHub CLI 2.48 or later. The target repository must live on github.com; GitHub Enterprise Server is not supported.
Security caveats operators should not skip
Both Harness presets used by the bundle run with danger-full-access as the operator's OS user. Account wrappers pick the agent's credentials but do not isolate the process; a worker can execute programs by path and read the operator's home directory. The deployment guide recommends a dedicated OS account that holds only the agent's credentials when isolation matters.
Architecture in brief
The architecture guide describes a Router preset for the main Session and worker presets for Executor sessions, coordinated by an Engine that owns Tickets and Assignments. Persistent state lives in one SQLite file, orchestrator.sqlite, opened only by the Store. That split—main Session for assessment and routing, Executor sessions for ticket work—is how the prototype separates orchestration from execution.
What it means
Analysis (ours). Publishing the harness bundle turns DeepSeek's super-agent coding experiment into something teams can replay on their own repos, with explicit guardrails on who can trigger work and where results land. It sits alongside DeepSeek's other public engineering, such as DeepSeek Prices Agent Web Search at $0.75 per 1,000 Calls, but addresses a different problem: long-running repository automation rather than priced API tools. Treat it as a reference implementation to study and harden, not as supported product software.
What to watch
Whether teams outside DeepSeek adopt the bundle on real repositories, how the maintainers evolve the published docs versus the internal design article, and any follow-up changes to Harness presets, allowed-list policy or deployment security guidance in the GitHub repository.