# DeepSeek open-sources dsh-superagent, a prototype coding agent that picks up work from GitHub on its own

> DeepSeek's Harness team published an MIT-licensed research prototype that watches a GitHub repo, takes authorized requests and answers with pull requests. It passed 11,000 stars within days — but DeepSeek itself says it is not production-ready and runs with full user access.

Published Nov 2, 2026 by TechMeasured Staff · Repos · Canonical: https://techmeasured.com/repos/deepseek-dsh-superagent-proactive-coding-agent-prototype

DeepSeek's Harness team has published **dsh-superagent**, an open-source prototype of a "proactive" coding agent that sits on a GitHub repository, waits for authorized people to ask for work, and replies with results such as pull requests. The repo went up on October 31, according to an independent write-up on DEV Community, and its GitHub page already showed about **11.5K stars and 1.2K forks** when we checked.

Why it matters: most coding agents today wait in a terminal or IDE for one person's prompt. This one is designed to live inside a team's workflow — listening to issues and comments, running several jobs in parallel, and only bothering humans when they actually need to decide something. The README sums up the goal as taking "as little human attention as possible."

## What it does

Per the README, dsh-superagent is packaged as a bundle for **DeepSeek Harness (dsh)**, the open-source agent harness DeepSeek released as a developer preview in August, which TechNode reported at the time. The superagent:

- **Listens to GitHub** and, optionally, **Feishu**.
- **Opens a work item** when an allowed person requests something, works on it in its own agent session and directory, and **replies in the original conversation**.
- Provides a **web panel** listing items that need a person's attention.
- **Acts only when asked** or when a rule the team wrote allows it.

## How it works

Dataconomy, which went through DeepSeek's accompanying design paper — titled *Exploring Super-agent Design: Towards a More Autonomous Agent System* — describes a router-and-worker split:

- A long-running **Router** handles conversation and dispatch with a smaller toolset.
- Each authorized request becomes a **Ticket** with a fixed "Mandate" recording why the work may proceed, and runs in an isolated **Worker** session.
- Later pull-request comments are routed to the right Worker without a fresh mention.
- Proactive behaviour has three modes — off, observe and auto — with **observe as the default**, meaning it logs what it would have picked up without acting.

Dataconomy also reports the bundle is TypeScript, at version 0.1.0.

## DeepSeek's own test numbers

DeepSeek's tests, as summarised by Dataconomy, used simulated chat and CI message streams of about 300 messages containing about 20 work requests, run five times per configuration:

| Measure | Result (DeepSeek's figures) |
|---|---|
| Median task completion | 12.4 minutes |
| 95th-percentile completion | 15.7 minutes |
| Single-session baseline, 95th percentile | 29.5 minutes |
| Router latency as events rose from ~1 to 40/min | 2.5 → 3.9 seconds per request |

These are company figures from **synthetic messages and simulated tools**; Dataconomy notes they don't measure real-world code quality.

## How to try it

Setup is agent-guided. The README tells you to open the repo with a coding agent and ask:

```text
Read SETUP.md in this repository and set up dsh-superagent with me.
```

For development, the README lists one command, which runs the typecheck, tests and build:

```bash
pnpm check
```

Manual setup, configuration and upgrades are covered in the repo's `docs/006-deployment.md`. Dataconomy lists the requirements as Node.js 22.19+ (or 24+), Git 2.32+, Python 3.8+, tmux and GitHub CLI 2.48+.

## Limitations and who should care

- **It's research code.** The README calls it a research prototype "for reference only," not an official DeepSeek product. It says much of the code was written by coding agents, may contain bugs, and differs from both DeepSeek's internal version and the design paper — some designs there aren't implemented.
- **Isolate it.** According to Dataconomy, DeepSeek warns the agent runs with full operating-system user access. Run it in a VM or container with a throwaway GitHub token, not on your main workstation.
- **Scope is narrow.** One repository at a time, with GitHub and Feishu as the only listed sources.
- **Permissions are on you.** The DEV Community reviewer suggests testing who can trigger work, how ambiguous requests are handled and what a wrong action costs before trusting it.

**Who should care:** teams building or evaluating agent workflows, and anyone curious how a frontier lab structures a multi-session coding agent. The numbered docs (philosophy, authorization, execution, prompts, architecture) are arguably the most valuable part right now. If you just want a dependable coding assistant, this isn't it — yet.


## Sources

1. [GitHub — deepseek-ai/dsh-superagent-prototype](https://github.com/deepseek-ai/dsh-superagent-prototype)
2. [Dataconomy — DeepSeek releases open-source dsh-superagent coding agent](https://dataconomy.com/2026/11/02/deepseek-releases-open-source-dsh-superagent-coding-agent/)
3. [DEV Community — DeepSeek Super-agent: what it actually is, and what it is not yet](https://dev.to/paullovelace/deepseek-super-agent-a-proactive-open-source-coding-agent-prototype-what-it-actually-is-and-438a)
4. [TechNode — DeepSeek opens its open-source Harness to developers](https://technode.com/?p=198079)