Back home

dsh-superagent

github 0.88
Visit source

Intro

dsh-superagent is an open-source proactive coding agent released by DeepSeek on October 31, 2026 as the repository deepseek-ai/dsh-superagent-prototype. It is packaged as a bundle for DeepSeek Harness (the "DSH" in its name), DeepSeek's agent runtime, and it accompanies the team's article Exploring Super-agent Design: Towards a More Autonomous Agent System.

Instead of waiting in a chat window, dsh-superagent "lives" on one GitHub repository. It listens to GitHub events and, optionally, Feishu (Lark) chats. When an allowed person asks it for work, it opens a work item, does the job in its own agent session and directory, and replies in the original conversation — typically with a pull request. A small web panel shows what needs a human.

DeepSeek is explicit that this is research code, not an official product: much of it was written by coding agents, it differs from their internal version, and some designs in the article are not implemented. It is MIT-licensed and passed 1,300 GitHub stars within hours of release.

Explain

How it works

  • Always-on, one repo: the agent is bound to a single GitHub repository and reacts to issues, PR comments and (optionally) Feishu messages.
  • Authorization first: it acts only when an allowed person asks, or when a rule the team wrote permits it. Proactive rules can start in a mode that only records what the agent would have done, before being trusted to act.
  • Work items and isolation: every request becomes a work item with its own agent session and working directory, so parallel tasks do not trample each other.
  • Answer where you asked: results go back to the original thread — a PR, a reply, or a question for a human.
  • Panel: a web panel lists items that are blocked on people.

Design philosophy

The core idea is "take as little human attention as possible, and give as much value as possible". DeepSeek argues single-task agents are already strong; the bottleneck is human attention. An agent that can act for a whole team can also waste a whole team's attention, so it speaks only when someone must act or can use the result.

Documentation

The repo ships numbered docs: 000 Philosophy, 001 Domain model, 002 Authorization, 003 Execution, 004 Prompts, 005 Panel, 006 Deployment, 007 Architecture and 008 Cookbook (adding a Source). pnpm check runs typecheck, tests and build.

Caveats

  • Research prototype; review the code before running it.
  • Requires DeepSeek Harness.
  • Not identical to DeepSeek's internal system.

Usage

Getting started

  1. Make sure your team already uses DeepSeek Harness.
  2. Clone https://github.com/deepseek-ai/dsh-superagent-prototype.
  3. Setup has many steps, so the README suggests letting a coding agent guide you: ask it "Read SETUP.md in this repository and set up dsh-superagent with me." It asks one question at a time, shows the full plan, then installs and configures everything.
  4. Configure the GitHub repository it watches, the allow-list of people who may assign work, and optionally the Feishu integration.
  5. Start with low-risk tasks — e.g. reproducing and fixing small bugs or updating docs — and review every PR it opens.

Good fits

  • Small teams that want an always-on helper on one repo.
  • Teams already chatting in Feishu who want bugs discussed in chat to turn into PRs.
  • Researchers studying multi-session, attention-aware agent design.

Links