DeepSeek releases MIT-licensed Super-agent prototype that auto-fixes GitHub bugs
AI
Reported by Dr. Kai Wei, Frontier AI Desk
- DeepSeek published a two-layer Super-agent prototype under MIT licence, handling GitHub events and auto-generating fixes.
- The Router dispatches tasks while Workers execute them, reducing inference time per decision from 6.4 s to 1.6 s in tests.
- Preliminary synthetic tests showed 95 % of 20 work requests completed within 15.7 minutes, highlighting scalability challenges.
DeepSeek’s Harness team unveiled an open-source Super-agent that routes work to isolated Workers, autonomously detects code errors and opens pull requests.
«DeepSeek Harness Team Open-Sources a Super-agent Prototype That Finds and Fixes Bugs on Its Own»
Summary and commentary by Aura-V. Facts, figures and opinions belong to the original.
DeepSeek’s Harness team announced a Super-agent prototype on October 31, releasing the code on GitHub under an MIT licence Pandaily. The system comprises a resident Router that receives human interactions across channels and delegates tasks to isolated Worker sessions, which execute code fixes and open pull requests without direct user contact.
The architecture is event-driven: chat messages, emails, review comments and CI results trigger deterministic filters and a model before reaching Workers. A Watchdog session monitors progress, and Workers subscribe only to relevant threads, allowing follow-up events to bypass the Router.
In synthetic workloads of roughly 300 messages and 20 work requests, the Super-agent completed half the tasks in 12.4 minutes and 95 % within 15.7 minutes, outperforming a single-session baseline despite higher cached token usage.
Our read: the prototype demonstrates a viable pattern for continuous, autonomous code maintenance, but practical adoption will depend on addressing the highlighted scalability and security concerns.
Interested in implementing this technology inside your company?
Our studio designs custom agent meshes, data pipelines, and outbound GTM systems for forward-looking enterprises.
