DeepSeek super-agent prototype: MIT coding agent that opens PRs
DeepSeek published dsh-superagent, an MIT-licensed proactive coding agent that watches a GitHub repo and answers with pull requests. Copy its rules, not its code.
DeepSeek put a proactive coding agent on GitHub under the MIT license. The repo, dsh-superagent-prototype, passed 6,500 stars by November 1. It listens to one GitHub repository, takes work when an allowed person asks, and answers in the same thread, often with a pull request. DeepSeek is clear that it is a research prototype and "not an official DeepSeek product." The code is not the interesting part. The rules around it are.
What actually happened
The repo appeared on the deepseek-ai GitHub org with one commit, last updated October 31. It is TypeScript, and it runs as a bundle on top of DeepSeek Harness, the MIT agent runtime DeepSeek open-sourced in August. Per the README:
- Inputs: GitHub events, and optionally Feishu, the chat app.
- Work model: each request becomes a work item, with its own agent session and its own directory.
- Output: a reply in the original conversation, for example a PR.
- Oversight: a web panel that shows what needs a person.
- Setup: you ask your own coding agent to read SETUP.md and walk you through install.
The README warns that much of the code was written by coding agents, may have bugs, and differs from DeepSeek's internal version. Some designs from DeepSeek's design article are future work, not shipped.
Why a proactive coding agent matters for your business
The philosophy doc is the product. The stated goal is to take as little human attention as possible and give as much value as possible. The agent acts only when asked or when a rule the team wrote allows it, and speaks only when someone must act. That is the right rule for any agent you put in a shared channel, coding or not. An agent that comments on everything trains your team to ignore it.
Authorization is a document, not a vibe. The docs give authorization its own chapter. Before you let any agent open PRs on your repo, write down who can assign it work, which branches it can touch, and what it must never merge. Then check that the tool enforces that list.
Don't run it in production as shipped. DeepSeek says so itself. Fork it, read it, and take the patterns: one work item per request, isolated directories, a human review queue. Those transfer to Claude Code, Codex or your own loop.
Key takeaways
- DeepSeek published dsh-superagent-prototype under MIT; it passed 6,500 GitHub stars by November 1
- It watches one GitHub repo, takes requests from allowed people, and replies with results such as PRs
- Each request gets its own work item, session and directory; a web panel flags what needs a human
- DeepSeek calls it a research prototype, not an official product or production software
- Copy the rules: act only when asked, speak only when someone must act, write authorization down
Want an agent that does repo work without spamming your team? We build coding and ops agents with written authorization rules, isolated workspaces and a human review queue, on whichever model you choose. See what we build or tell us which repo eats your week.
Sources: DeepSeek on GitHub: dsh-superagent-prototype, DeepSeek GitHub org.
- #deepseek
- #ai-agents
- #open-source
- #coding-agents
- #github
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Vercel Labs agent-desktop 1.0: computer use without pixels
agent-desktop 1.0 is an Apache-2.0 CLI under Vercel Labs that lets AI agents drive macOS, Windows and Linux apps through the accessibility tree. Here's where it fits.
Read itGoogle Jules shutdown: read-only Nov 11, gone Dec 31
Google is shutting down Jules, its async coding agent. Sessions end November 11 and the service closes December 31. Here's what to export and where to move.
Read it