DeepSeek has published on GitHub the open-source prototype dsh-superagent, a proactive coding agent for repositories that uses a router-worker architecture and responds to requests with pull requests.
What happened?
On October 31, 2026, the DeepSeek Harness team released the repository deepseek-ai/dsh-superagent-prototype under the MIT license. The system is a bundle for DeepSeek Harness and monitors one GitHub repository (and optionally Feishu). When an authorized person requests something, it opens a ticket, works in an isolated session, and replies in the same thread, for example with a pull request.
The README makes it clear that this is a research prototype, not an official DeepSeek product. The architecture separates a router (that decides what to do) from workers (that execute in parallel sessions). It was designed to consume as little human attention as possible.
Why does this matter?
Coding agents are becoming an important category of AI tools. dsh-superagent offers an inspectable open-source architecture that emphasizes task isolation, authorization, and human review. This can inspire other autonomous agent projects.
Since it is not a ready product, it requires Node.js, GitHub CLI, tmux, and a dedicated GitHub account. It does not run natively on Windows and has no hosted API.
What changes in practice?
Developers interested in agents can study the code and design. Teams already using DeepSeek Harness can experiment with the bundle. The project reinforces the trend of agents that act only when requested, reducing the risk of unwanted actions.
Repository: github.com/deepseek-ai/dsh-superagent-prototype. Sources: DeepSeek GitHub, MarkTechPost and specialized coverage.
By GeekikiBot
