How to set up the DeepSeek super-agent, step by step

To set up the DeepSeek super-agent, you install DeepSeek Harness, give the agent its own GitHub account, answer a short list of questions about your repository, and start the service in observe mode. This guide follows the setup guide in DeepSeek's repository, read on November 1, 2026. It is for developers who want to try the prototype on a throwaway repository. We did not run it ourselves, so every step below comes from DeepSeek's docs.

DeepSeek calls the project a research prototype, "not an official DeepSeek product", so expect rough edges. For what it is and how it decides what to do, see our news article on the release.

What you need first

The setup guide lists these requirements:

• Node.js 22.19 or newer, or Node.js 24 or newer
• pnpm
• git 2.32 or newer
• tmux
• Python 3.8 or newer
• GitHub CLI 2.48 or newer
• DeepSeek Harness (DSH), in the version range named in the repository's package.json
• A model provider configured for DSH. The setup page does not say which model to use, and configuring one is a manual step for you.
• A GitHub account for the agent, separate from yours, so its commits and comments are attributed to it

A Feishu bot is optional. Feishu is the chat app from ByteDance, and the agent can listen to it as well as to GitHub.

How to set up the DeepSeek super-agent

The repository's SETUP.md is written so that a coding agent can walk you through it. You can also do the steps by hand, and the manual route is in docs/006-deployment.md. The flow is:

1. Run the one-time probe. It checks which tools and DSH profiles you already have.
2. Answer the questions, one at a time: where requests come from, which repository the agent works on, the agent's GitHub account, which GitHub users are allowed to ask it for work, Feishu settings if you use them, where the checkout lives, the commit identity, and whether auto-merge is on.
3. Review the single plan it produces and confirm it.
4. Do the operator-only steps yourself in your own terminal: create the agent's GitHub account, sign in to the CLIs, and set up Feishu if you use it. DeepSeek says secrets are never requested in chat.
5. Install and start the service, then save the settings through its management API.
6. Check it with the status script. The guide suggests an optional hourly repair check, and you should read the operating guidance it points to.

Start in observe mode

The default accept mode is observe. In this mode the agent still answers explicit requests from people on your allowed list, but anything it infers on its own, such as a bot message or a follow-up from an unlisted person, is only recorded as something it would have accepted. Nothing inferred is acted on.

DeepSeek's authorization doc suggests this order: run in observe, review the recorded decisions, then switch to auto once you trust the rules you wrote. Keep the allowed list short, limited to yourself, while you learn how it behaves. Auto-merge is off by default. If you turn it on, it needs a separate account that is allowed to merge.

Common problems

• The agent does nothing. Check that the person asking is on the allowed list. Requests from bots are never explicit, and an unlisted person's request is only inferred.
• Nothing inferred is accepted. That is observe mode working as intended. Switch to auto only after reviewing its recorded decisions.
• Setup stalls at the model. The guide treats model-provider setup as a manual operator step.
• Checking your changes to the code. The repository uses pnpm check, which runs the typecheck, the tests and the build.

Before you give it real work

An agent that reads messages and issues from outside is open to prompt injection, where text in an issue tries to steer it. DeepSeek's docs say messages, comments and quoted documents cannot change permissions, but the README also warns the code may contain bugs and should be reviewed before use. Use a repository with nothing private in it first.