DeepSeek’s New Open-Source Coding Agent Writes Pull Requests for You

Most coding agents wait for you to ask. DeepSeek’s new one doesn’t. On October 31, the Chinese AI lab open-sourced dsh-superagent — a prototype coding agent that watches a team’s GitHub repository, picks up requests, and answers with finished work like pull requests. It collected thousands of stars within a day.

So what makes this one different from Claude Code or Codex?

A coding agent for the whole team

Here’s the idea: instead of opening a chat session and assigning a task, the agent lives inside your repository. It’s built as a bundle for DeepSeek Harness, the lab’s open agent framework. It listens to one GitHub repo — and optionally to Feishu, the workplace chat app — and when an approved team member asks for work, it opens a work item, does the job in its own session and folder, then replies right where the request was made. The classic reply? A pull request, ready for review.

A local web panel shows only the items that actually need a human decision. The design rule is stated plainly in the project docs: take as little human attention as possible, and give as much value as possible.

Under the hood: routers, workers, and a watchdog

The architecture splits the agent in two. A persistent “Router” talks to people and keeps shared context, while separate “Worker” sessions run tasks in parallel — so the agent can take a new request before the last one finishes. A Worker that owns a pull request follows its comments directly, and a Watchdog process monitors for failures and stalled loops.

In DeepSeek’s own simulated tests, the time to finish the slowest 5% of tasks fell from 29.5 minutes to 15.7 minutes. Take that number with a pinch of salt — it’s the lab’s own simulation, not an independent benchmark.

What you’ll need to run it

This is a research prototype, not a polished product, and the setup shows it. You need a dedicated GitHub user account for the agent, Node.js 22.19 or later, GitHub CLI 2.48 or later, and a model provider configured in DeepSeek Harness. Only github.com works — GitHub Enterprise Server isn’t supported. And note this: the agent’s sessions run with full access as your OS user, so DeepSeek recommends a separate OS account for isolation.

DeepSeek is unusually blunt about the limits. The README calls it “a research prototype for reference only, not an official DeepSeek product,” warns the code may still contain bugs, and even admits much of the engineering was done by coding agents themselves.

Why it matters

Bloomberg framed the release as a direct shot at Anthropic’s Claude Code — and it’s easy to see why. While most agents are personal assistants, DeepSeek is sketching something more ambitious: an agent that acts like a tireless junior teammate, quietly fixing bugs and answering code reviews in the background. The prototype is narrower than that vision, but as reference architecture for where coding agents are headed, it’s worth studying.

Also read: GitHub’s Octoverse 2026: AI agents are writing the pull requests now · 5 new AI tools worth trying this month

Outbound sources: Skimwire on the dsh-superagent release · Complete AI Training’s breakdown