DeepSeek put out a prototype on Saturday that tries a different answer to Claude Code:
dsh-superagent. The Harness team announced it in a WeChat post, and Bloomberg reported it
first. The repository has been public since Saturday morning and passed 9,000 stars by
Sunday evening.
The agent attaches to exactly one GitHub repository and listens there, optionally in Feishu too. When someone with the right permissions asks for work, it opens a work item, handles it in its own agent session and its own directory, and answers back in the original conversation, usually with a pull request. A web panel shows what still needs a person.
The team is unusually blunt about what this isn’t
The warning sits at the top of the README: research code for reference, not an official DeepSeek product, not production-ready, probably still buggy, inconsistent in style and engineering practice. It grew out of the team’s day-to-day work, with “much of the engineering done by coding agents.”
There is a second caveat. The published version is neither the internal one nor the one described in the paper “Exploring Super-agent Design: Towards a More Autonomous Agent System.” Some designs in that paper are directions for later and aren’t implemented here.
Technically it’s a bundle for DeepSeek Harness, MIT licensed, and setup is handled by — of
course — a coding agent: you ask yours to read SETUP.md and walk the installation with
you.
The real idea is a rule about attention
The sentence that separates this from a weekend repository sits in the philosophy file: take as little human attention as possible, and give as much value as possible. Two constraints follow. The agent acts only when asked, or when a rule the team itself wrote allows it. And it speaks only when someone has to act or can use a result.
That is the inverse of what proactive agents normally do. An agent that can work for a whole team can also burn a whole team’s attention, and the rule is aimed squarely at that.
The comparison with Anthropic’s own line is the interesting part. A week ago Anthropic published its thinking on when subagents belong in the main loop, and the argument there runs on context and tokens. DeepSeek’s argument runs on the patience of the people in the chat. Both questions are fair. Hardly anyone is asking the second one.