DeepSeek open-sources DSH Superagent coding agent on GitHub
The research prototype splits work between a Router and parallel Workers and cut task time by about 47% in a simulated test, DeepSeek said.
DeepSeek published a technical document on its super agent design and released the source code of DSH Superagent (dsh-superagent) on GitHub, according to aitimes.com. The document, titled "Exploring Super Agent Design: Toward More Autonomous Agent Systems," was posted through the company's official WeChat account on October 31.
The architecture divides tasks between a "Router," which reads developer team conversations and GitHub events to allocate work, and "Workers," which fix code and resolve errors in parallel inside isolated environments. When an error comes up in a developer team chat, the agent can analyze the problem and begin fixing it under pre-approved rules, then submit a pull request on GitHub for review.
DeepSeek said its internally operated super agent also takes part in real bug fixes and responses to code reviews. In a simulated test involving about 300 messages and 20 task requests, the system completed 95% of tasks within 15.7 minutes, compared with 29.5 minutes for the existing single-session method, a reduction of about 47%.
The company labeled the GitHub code a research prototype rather than an official product and said the public version differs from its internal system. The public release is designed to automate multiple coding tasks for a single software project, and some automation features are not yet implemented. The code is published under the MIT license, which permits commercial use, and is built to connect various AI models.
Team formed earlier in 2026
DeepSeek formed a dedicated harness team in May 2026 with Anthropic's Claude Code as its target competitor, and expanded hiring of development staff in June 2026. In August 2026 the company released the open-source development environment DeepSeek Harness.
Quick answers
What is DSH Superagent?
It is DeepSeek's super agent architecture that splits coding work between a Router and parallel Workers. DeepSeek released its source code on GitHub as a research prototype.
Can the code be used commercially?
Yes. DeepSeek released the code under the MIT license, which permits commercial use.
How fast is it compared with the existing method?
In a simulated test with about 300 messages and 20 task requests, DeepSeek said the super agent completed 95% of tasks within 15.7 minutes, versus 29.5 minutes for the existing single-session method.