DeepSeek Super-agent: an open source proactive coding agent

A research prototype that listens to a GitHub repository and team chats, then opens pull requests when asked.

DeepSeek released Super-agent on October 31, 2026, an open source prototype of a coding agent that watches one GitHub repository and, optionally, Feishu team chats. It takes requests from approved people and answers with work such as pull requests. DeepSeek says it is reference code, not an official product.

Key facts

DeepSeek published the dsh-superagent-prototype repository on GitHub on October 31, 2026, under the MIT license.
Super-agent is packaged as a bundle for DeepSeek Harness, the company’s open source agent framework.
It listens to one GitHub repository and, as an option, Feishu (Lark) group chats, and works on requests from people who have permission.
DeepSeek calls it a research prototype “for reference only” and says it is not production ready.
The repository passed 4,000 stars within about a day, according to ShiftDelete.

What happened

DeepSeek’s Harness team announced Super-agent in a WeChat post on Saturday, October 31, and published the code on GitHub the same day. The README describes it as a proactive coding agent for a single GitHub repository. When someone with permission asks for work, the agent opens a work item, carries it out in its own session and folder, and replies in the conversation where the request started, for example with a pull request. A local web panel shows only the items that need a human decision.

ShiftDelete and Startup Fortune describe the design in more detail. A persistent Router talks with people and keeps shared context, while separate Worker sessions run tasks in parallel. A Worker that owns a pull request follows the comments on it, and a Watchdog process looks for failures and stalled loops. Work that nobody asked for needs a rule written by the team, and ShiftDelete reports that this kind of work starts in a mode that only records what the agent would have done.

DeepSeek also published a design article with the code. The README warns that the public version differs from the one the team uses internally and that some ideas in the article are not built yet. Bloomberg framed the release as a challenge to Anthropic’s Claude Code.

Why it matters

Most coding assistants wait for you to open them and type a request. Super-agent tries a different shape: it sits next to the places where a team already talks, picks up requests from those threads and returns results there. The stated goal in the repository is to take as little human attention as possible while giving as much value as possible.

Because the code is MIT licensed, anyone can read how the routing, permissions and parallel sessions are put together, and pick their own model provider through Harness. For small teams and solo developers this is a useful blueprint to study, even if they never run it.

The limits matter just as much. DeepSeek says sessions run with full access as the operator’s user account and recommends a separate account for isolation. Startup Fortune notes that Harness has not had a security audit. After this week’s fake Claude Code extension in the Cursor marketplace, it is a good reminder to treat any agent with repository access as something to test carefully, away from real work.

How to try it

If you are a developer, the README suggests a guided setup: you ask a coding agent to read SETUP.md and it walks you through one question at a time. ShiftDelete lists the requirements: a dedicated GitHub account for the agent, Node.js 22.19 or later, GitHub CLI 2.48 or later and a model provider set up in DeepSeek Harness. Only github.com is supported. Use a test repository and a separate user account on your computer.

For everyone else, the useful idea is deciding which tasks to hand to an agent and which to keep. Our free Plan my week prompt helps you lay that out. Fill in the brackets like this:

Fixed commitments: “Mon to Fri 10:00 stand up, Wed 14:00 client review”
Goals for the week: “Fix the login bug, review two pull requests, try an AI coding agent on a test repository”
Free hours per day: “About 2 hours of focused time”
When you work best: “Morning”

Check that the plan gives you one clear priority per day, leaves buffer time for testing the agent, and keeps the review of its pull requests on your own schedule. The prompt plans your time; it does not install or run any software.

Try it

Plan my week

Frequently asked questions

Is DeepSeek Super-agent free to use?

The code is published under the MIT license, so you can read, change and self host it. You still need a model provider configured in DeepSeek Harness, and that provider may charge for usage. DeepSeek describes the release as reference code, not a supported product.

Can Super-agent work on its own without being asked?

By default it acts when a person with permission asks. Unrequested work needs a rule that the team writes, and ShiftDelete reports that such work starts in a mode that only records what the agent would have done, so people can review it first.

Is DeepSeek Super-agent safe to run on my main computer?

DeepSeek says it is not production ready and that sessions run with full access as your user account. It recommends a separate operating system account. Testing it on a throwaway repository is the sensible first step.

Sources