Skip to content
Neural Daily

ToolsNews

DeepSeek dsh-superagent: a proactive coding agent

🇭🇺 Olvasd el magyarul →

The DeepSeek Harness team has released a dsh-superagent prototype under the MIT license. It watches one GitHub repo, runs requests in parallel isolated sessions, and finished tasks about twice as fast as a single-session agent in simulation.

By: Mikola Máté Patrik · · 2 min read

GitHub page of the deepseek-ai/dsh-superagent-prototype repository
Image: DeepSeek · Screenshot · Screenshot of the official DeepSeek GitHub repository · Source

What it is

dsh-superagent is a TypeScript bundle for DeepSeek Harness. Version 0.1.0 landed as a single commit on GitHub. It watches one repository and, optionally, Feishu chat. When an allowed person addresses it, it opens a work item called a Ticket that runs in its own session and directory. Results usually come back to the original thread as a pull request.

The README states the core rule plainly: take as little human attention as possible and give as much value as possible. The agent confirms once, stays quiet while working, and writes again only with a result, a question or a failure.

Router and Worker

DeepSeek argues that today's agents treat the session as the basic unit, while a person can only follow 3 to 5 sessions at once. So the system has two layers:

  • Router: a long-running main session that evaluates incoming events and dispatches work.
  • Worker: isolated sessions, each carrying one Ticket. Later PR comments reach the right Worker without a fresh mention.
  • Proactive work: observe mode by default, where the agent only logs what it would have accepted. Teams must switch on auto mode themselves.

What was measured

The test used simulated group chat and GitHub CI streams with about 300 messages per run, of which about 20 were real work requests. The superagent reached a p95 task time of 15.7 minutes, versus 29.5 minutes for a single-session baseline with subagents. Under a neutral prompt, the baseline did not call a subagent once in five runs (MarkTechPost).

The team spells out the limits: all figures are self-reported, using synthetic messages and simulated tools. Real coding quality was not measured.

What to watch

The agent runs with full OS-user permissions, so isolate it in a separate user or VM. It needs Node.js 22.19+ or 24+, git, tmux and the GitHub CLI. This is a research prototype, not a product. The Router-Worker pattern still shows where agentic development is heading: the real bottleneck is not the model but how many parallel threads one person can follow.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles