dsh-superagent-prototype: A coding agent that automatically handles everything from GitHub listening to PRs

deepseek-ai/dsh-superagent-prototype

·2026.11.01 00:40

This is a proactive coding agent that operates on a single GitHub repository. It is packaged and distributed as a DeepSeek Harness bundle, monitoring GitHub events and, optionally, Feishu messages. When an authorized user requests a task, the agent creates a work item, writes code in an independent session and directory, and then responds to the original conversation with the deliverable (e.g., a Pull Request).

The core design philosophy is 'minimal human attention, maximum value.' To avoid wasting the team's attention, the agent acts only on explicit requests or rules written by the team. It also sends messages only when someone can take action or utilize the results. A web panel allows you to visually check items currently requiring human intervention.

This project is a research prototype, not an official DeepSeek product. The code was largely written by a coding agent during internal development, and not all features mentioned in the 'Exploring Super-agent Design' paper are implemented. It is released under the MIT license, but it is intended to be used as a reference rather than applied directly to production environments; code review and environment adaptation are required before use.

This is suitable for developers looking to build autonomous agent systems or redesign the operational methods of existing coding agents. In particular, teams aiming to improve team-level development efficiency by combining GitHub workflows with agent automation can find it useful. The documentation sequentially explains the domain model, permission management, execution structure, prompt design, and deployment methods, making it easy to understand step by step.

GitHub
GitHub repository

deepseek-ai/dsh-superagent-prototype

Towards a More Autonomous Agent System

TypeScript

This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.