DeepSeek unveils open-source super-agent framework for GitHub automation

DeepSeek has open-sourced dsh-superagent, a research prototype designed as a proactive coding agent for GitHub repositories. The tool, released under an MIT license, operates as part of the DeepSeek Harness system and integrates with GitHub and Feishu. It processes allowed requests into parallel work items, each running in its own session. The release includes a design post titled 'Exploring Super-agent Design: Towards a More Autonomous Agent System'. The agent follows a philosophy of minimizing human intervention while maximizing output. It confirms once, then works silently, and only communicates results, questions, or failures. DeepSeek argues that mainstream agents are limited by human attention, as users can only actively follow 3 to 5 sessions at once. The prototype features a Router for interaction and a Worker layer for execution, with Router decisions increasing by 30% and cache-read tokens rising by 37% in tests. The system supports three authorization modes and includes ideas for future enhancements like an external Watchdog and asynchronous ReAct loops. DeepSeek tested the prototype using simulated group chats and GitHub CI streams, with each run involving about 300 messages, 20 of which were real work requests.