--- title: "DeepSeek Open-Sources Super-agent, a Team Coding Agent" description: "DeepSeek put Super-agent, a prototype coding agent that watches one GitHub repo and optional Feishu chats, on GitHub under MIT. Here's what it needs to run." url: https://publishmyblog.com/ai/deepseek-super-agent-open-source/ section: "AI" author: "Saeed Shahriar" published: 2026-11-02T06:53:01.035Z --- # DeepSeek Open-Sources Super-agent, a Team Coding Agent It's a research prototype, and its agent sessions run with full access as your own OS user. DeepSeek has released Super-agent, a coding agent that watches one GitHub repository for your team and picks up work when someone asks. **The code is on GitHub under the MIT license.** So should you put it on your product's repo this week? **Probably not yet**, and DeepSeek says much the same. ## What Super-agent does The [dsh-superagent repository](https://github.com/deepseek-ai/dsh-superagent-prototype) calls it **a proactive coding agent for one GitHub repository**. It's a bundle for DeepSeek Harness. It listens to GitHub and, if your team works there, to Feishu group chats as well, which The Next Web describes as ByteDance's workplace chat app. Someone on your allowed list asks for work, and **the agent opens a work item in its own session and directory**, then answers in that same thread, for example with a pull request. A web panel shows you only what needs a person. [ShiftDelete's report](https://shiftdelete.net/en/deepseek-coding-agent) splits it into three parts: - a "Router" that talks to people and keeps the shared context - "Worker" sessions that run tasks side by side, so a new request doesn't wait for the last one to finish - a Watchdog that looks for failures and stalled loops Want it to act on its own? **You need a rule your team writes.** Even then, ShiftDelete says, it starts by just recording what the agent would have done. ## Who built it, and what DeepSeek warns DeepSeek's Harness team described the design in a WeChat post on October 31, [The Next Web reports](https://thenextweb.com/news/deepseek-super-agent-open-source-coding-agent). **Bloomberg says it's aimed at Claude Code.** The README calls it a research prototype, "not an official DeepSeek product." It says coding agents did much of the engineering, so **it may still contain bugs**. Then there's a gap: the public code isn't identical to the version DeepSeek's team uses internally. Still, interest came fast. ShiftDelete says the repo passed 4,000 stars within a day of going public. ## What you need to run it Setup has so many steps that the README tells you to **ask your own coding agent to read SETUP.md** and set the whole thing up with you, one question at a time. According to ShiftDelete, you'll need: - a dedicated GitHub user account for the agent - Node.js 22.19 or later and GitHub CLI 2.48 or later - a model provider configured in DeepSeek Harness - a repository on github.com (GitHub Enterprise Server isn't supported) The fine print matters most. **Its sessions run with full access as the operator's OS user**, and DeepSeek recommends a separate OS account for isolation. ## Why it matters Running a small SaaS or a site with real code? **This is a coding agent that lives in your repo and team chat** and picks up work there, instead of waiting in a tool you open. Anthropic's [Claude Duet preview, an agent that messages you first](https://publishmyblog.com/ai/anthropic-claude-duet-agent-preview/), points the same way. I like the rule. The README asks the agent to **spend as little of your team's attention as it can** while giving back as much value as it can. That's a fair test. But read the setup twice. **An agent with full access as your OS user is what you'd wall off.** We've seen [a backdoored SEO plugin release](https://publishmyblog.com/seo/rank-math-pro-backdoor/) reach sites in hours. My read, if you want to try it: - Put it on a spare machine or a separate OS account. - Give it a test repo and a fresh GitHub account, never your main login or the repo that pays your bills. - Write one small rule and watch what it says it would have done before you let it act. There's a team question too. TNW cites a January CooperBench study in which **coding agents working as a team had success rates about 30% lower** on average than each did alone. It didn't test DeepSeek's agent. Anyway, in the US public sector? TNW says **federal civilian agencies were told to remove DeepSeek** last month. Check first. ## Common Questions **What is DeepSeek Super-agent and what license is it under?** Super-agent (dsh-superagent) is DeepSeek's open-source research prototype of a coding agent for one GitHub repository, with optional Feishu chats. DeepSeek published the code on GitHub under the MIT license. **What do I need to run DeepSeek Super-agent on my own repository?** A separate GitHub user account for the agent, Node.js 22.19 or later, GitHub CLI 2.48 or later, a model provider set up in DeepSeek Harness, and a repository hosted on github.com rather than GitHub Enterprise Server. **Is DeepSeek Super-agent ready for production use?** No. DeepSeek calls it a research prototype for reference only, says it may contain bugs, and tells users to review and adapt the code before running it in their own environment. ## Sources - [deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent System](https://github.com/deepseek-ai/dsh-superagent-prototype), GitHub - [DeepSeek open-sources Super-Agent, a coding agent that reads team chats](https://thenextweb.com/news/deepseek-super-agent-open-source-coding-agent), The Next Web - [DeepSeek Coding Agent Super-agent Takes Aim at Claude Code](https://shiftdelete.net/en/deepseek-coding-agent), ShiftDelete