DeepSeek Harness team releases open-source Super-agent prototype
The DeepSeek Harness team has published the underlying architecture for their Super-agent as an open-source prototype on GitHub. The release is explicitly labeled as a research prototype for reference only, not an official DeepSeek product. The code originated from the team's internal development work, with much of the engineering performed by coding agents. The team warns that the software is…
Key points
- DeepSeek Harness team released dsh-superagent as an open-source research prototype on GitHub.
- The agent listens to GitHub and Feishu, handling work items and responding with pull requests.
- The code is not production-ready and differs from the internal version described in their research paper.
The prototype, named dsh-superagent, functions as a proactive coding agent for a single GitHub repository. It operates as a bundle for DeepSeek Harness and listens to GitHub and, optionally, Feishu. When an authorized user requests work, the agent opens a work item, processes it in its own session, and responds in the original conversation, often with a pull request. A web panel displays items requiring human attention. The design philosophy aims to minimize human attention while maximizing value, ensuring the agent acts only when asked or when permitted by team-written rules.
The published code is adapted for public release and is not identical to the internal version or the implementation described in the paper "Exploring Super-agent Design: Towards a More Autonomous Agent System." Some designs mentioned in the paper are future directions and are not implemented in this prototype. Setup is complex, so the repository includes a SETUP.md file designed to guide a coding agent through the installation and configuration process step-by-step.
DeepSeek Harness team published the underlying architecture for their Super-agent, an always-on autonomous agent, as an open-source prototype on GitHub.
github.com · 31 October 2026Loading the full article…
This text was published by github.com. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Developer outlines limits of testing Claude Code iOS apps in Mac simulator · 1 src
- Datasette-llm 0.1a9 auto-detects default LLM model · 1 src
- Claude Code burns 15–20% of tokens on a 3-word status line · 2 src
- Amazon CEO Andy Jassy says external AI shopping agents like Meta's Muse are still not great nor more convenient than shopping directly · 7 src
- Claude automates Google Sites build with AI-to-AI instructions · 7 src
Comments
via GitHub Discussions