DeepSeek superagent: an AI agent that opens pull requests by itself

DeepSeek has published dsh-superagent on GitHub. The DeepSeek superagent is a prototype AI agent that looks after a single GitHub repository: it takes requests from the team, works on them on its own, and replies in the same thread, for example with a finished pull request. The code is MIT-licensed, and according to ShiftDelete the DeepSeek Harness team announced it on WeChat on October 31.

What’s new here for AI?

Most coding agents we use every day wait for someone to type something into a terminal. DeepSeek superagent works the other way round: it’s a so-called “proactive” agent that listens to the repository (and, optionally, to chats in Feishu, aka Lark) and picks up work by itself when someone on the allowlist asks for it.

According to the README, each request gets its own task, its own agent session and its own directory, and the result goes back to wherever the request came from. On top of that there’s a web dashboard that shows only what needs a human decision.

Interestingly, the authors write that the code grew out of their team’s daily work, and that a large share of the engineering was done by coding agents (an agent built by agents ;) ).

The repository is described as “Towards a More Autonomous Agent System” and has more than 5.8k stars and 544 forks (as of November 1). ShiftDelete reports that it passed 4k stars within a day.

How does DeepSeek superagent work?

It isn’t a standalone app but a bundle for DeepSeek Harness, DeepSeek’s MIT-licensed agent harness, which is still in developer preview. The architecture docs describe this split of roles:

1. The main session (router) assesses incoming requests and decides whether to accept a task, hold it, or ignore it.
2. Worker sessions get one task each, every one with its own directory. They have no tools for asking people questions, and once the task is finished their calls are rejected.
3. A pull request is pinned to the task, and the task only ends when the PR is merged or closed, not when the agent reports that it’s done.

The most interesting part is the project’s philosophy. In the design principles document the rule is: take as little of people’s attention as possible and give as much value as possible. The agent speaks up only when someone has to do something or can use the result.

Can the agent act without asking?

It can, but not right away. The authorization document describes three modes for work nobody explicitly asked for:

1. off: the agent doesn’t consider anything it wasn’t asked to do.
2. observe (default): the agent assesses such requests but only records that it “would have accepted” the task, and does nothing.
3. auto: the agent accepts tasks on its own when they match rules the team wrote down in advance.

Direct requests from people on the allowlist work in every mode. Allowlisted bots can at most suggest candidates, and the agent itself can’t widen its own permissions: messages don’t change the allowlists, the rules or the scope of tasks.

What should you watch out for?

First, the project’s status. The authors say plainly that this is a “reference-only” research prototype, not an official DeepSeek product. The code may contain bugs, isn’t production-ready, and differs from the version the team uses internally.

Second, security. According to the deployment docs, both presets run in danger-full-access mode as your system user, and a worker can read your home directory. For isolation, the authors recommend a separate system account and a separate GitHub account for the agent (more on these risks in the post on AI agent security).

Third, the requirements. You need, among other things, Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, GitHub CLI 2.48+, pnpm, tmux, and a model provider configured in DeepSeek Harness. It works only with github.com; GitHub Enterprise Server isn’t supported.

How to try DeepSeek superagent?

Setup has quite a few steps, so according to the README it’s done by… an agent. Just ask your coding agent to read SETUP.md from the repository and walk through the configuration step by step (if you’re new to these tools, the Claude Code tutorial will help).

Once it’s running, open an issue or add a comment from an allowlisted account that mentions @<agent login> with your request. The task should show up in the dashboard within one polling cycle, a minute by default. The model itself is plugged in as a provider in DeepSeek Harness (DeepSeek V4.1 Flash was, by the way, recently free in OpenCode).

FAQ

Is DeepSeek superagent free?

The code is open source under the MIT license. You only pay for the model the agent uses (and for the machine it runs on).

Is it an official DeepSeek product?

No. The repository is on the deepseek-ai account, but the authors describe it as a research prototype, not an official product.

Does it work with GitLab or GitHub Enterprise?

According to the docs, only github.com is supported, and GitHub Enterprise Server isn’t. Optionally, the agent also listens to chats in Feishu (Lark).