DeepSeek Open-sources Super-agent Architecture Supporting AI Autonomous Bug Discovery and PR Submission

Techub News reports that the DeepSeek team has published a technical article open-sourcing the underlying architecture concept of their Personal Agent called "Super-agent." This architecture aims to free AI from single-session limitations, achieving cross-platform, multi-task parallel processing, and continuous online operation. In simulated multitasking environments, the median task completion time of the Super-agent is 12.4 minutes, outperforming traditional single-session solutions. The article elaborates on four core features of the Super-agent: continuous, efficient, proactive, and persistent. It adopts a two-layer architecture, with the upper layer Router responsible for receiving messages and scheduling tasks, and the lower layer Worker responsible for specific execution. It also designed subscription and direct-pass mechanisms to improve efficiency. The team introduced an asynchronous ReAct Loop, a Watchdog supervision mechanism, and solutions separating computation, state, and execution environments to support long-term autonomous AI operation. The DeepSeek team stated that this architecture allows AI to proactively discover bugs, locate issues, and submit fix PRs without human instructions while continuously monitoring workgroup messages and processing multiple complex tasks in parallel. Related prototype code has been open-sourced on GitHub. (Quantum Bit)