DeepSeek's answer to agent noise: speak only when it matters

What it does

dsh-superagent is a coding agent scoped to a single GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub (and optionally Feishu), and when an authorized person asks for work, it opens a work item, grinds on it in its own isolated agent session and directory, then replies in the original conversation — often as a pull request. A web panel surfaces whatever genuinely needs a human.

The interesting bit

The design principle is inverted from most agent hype: take as little human attention as possible, give as much value as possible. An agent that can act for a whole team can also waste a whole team’s attention, so it only acts when asked or when a team-written rule permits, and only speaks when someone must act or can use a result. That restraint is the feature.

Key highlights

• One agent session and working directory per work item — isolation by default
• Authorization model: only allowed people can trigger it, and only explicit rules expand its initiative
• Web panel for the human-decision queue
• Setup is itself agent-guided — it walks you through SETUP.md one question at a time
• Numbered docs (000–008) build a philosophy → domain → execution → deployment arc rather than a feature dump

Caveats

• Explicitly a research prototype, not an official DeepSeek product; expect bugs and inconsistent style
• Much of the code was written by coding agents, and the published version differs from both the internal one and the design article — some described designs aren’t implemented here at all

Verdict

Worth reading for anyone designing agent systems where the bottleneck is human attention, not model capability. Don’t deploy it near anything you care about without a careful review pass.