DeepSeek's Harness team open-sourced dsh-superagent, a prototype always-on coding agent

The DeepSeek Harness team released dsh-superagent, an MIT-licensed research prototype of a proactive coding agent for a single GitHub repository, according to MarkTechPost. It shipped as v0.1.0 alongside a design post titled Exploring Super-agent Design: Towards a More Autonomous Agent System. The team says it is not an official DeepSeek product or production-ready software.

The idea is that human attention, not model capability, is the bottleneck. The team argues that mainstream agents treat one session as the basic unit, and that a person can follow only three to five sessions at once. The design uses a long-running Router that handles interaction and dispatch, and isolated Worker sessions that do the actual work, each Ticket in its own session and directory.

In the prototype, the agent listens to GitHub and optionally Feishu, and turns allowed requests into parallel work items. Results return to the original conversation, often as a pull request. MarkTechPost lists three accept modes for proactive work: off, observe, which is the default and only records what it would have accepted, and auto. It needs Node.js 22.19 or newer or 24 or newer, git 2.32 or newer, Python 3.8 or newer, tmux and GitHub CLI 2.48 or newer.

DeepSeek reports a simulation with about 300 messages per run, around 20 of them real work requests, run 5 times per scheme. The super-agent had a p50 of 12.4 minutes and a p95 of 15.7 minutes, against a p95 of 29.5 minutes for a single session with subagents. MarkTechPost notes these are self-reported figures from synthetic messages and simulated tools, so real coding quality was not measured.

The Global Times reported that DeepSeek says an internal super-agent is running, helping reproduce and fix software problems, respond to code reviews and improve Harness itself. The published version differs from the internal one, and some features are still under development, per the project's GitHub repository.

For developers, the prototype is worth reading for its Router and Worker split and its gating of unasked work. MarkTechPost warns it runs with full OS-user access, so isolate it before trying it. Start in observe mode and watch the Tickets it would have accepted.

Building with AI? See this week's hot AI repos and how to use each one.