DeepSeek shares a research prototype for repository work

DeepSeek’s repository presents dsh-superagent as research code, not an official product or production-ready software. It serves one GitHub repository. When an allowed person requests work, the agent opens a separate session and directory, then responds in the original conversation, potentially with a proposed code change. The authors warn that bugs may remain and users should review and adapt the code.

The authorization document separates explicit requests from work inferred from messages. In the default observation mode, inferred candidates are assessed but never accepted: a positive assessment only records what would have been accepted. Explicit requests still work in every mode. Automatic acceptance of inferred work requires the automatic mode and a matching rule. Receiving a subscribed conversation as context does not itself grant new authority; the existing work remains bounded by its authorization.

The design principles address the human cost of supervision. The agent should acknowledge a request once, remain quiet while working and return with a result, a needed decision or a failure. Operators are told to begin with a short permission list and recorded proactive work, widening scope after examining the proposed actions. They remain responsible for messages posted under their accounts. These are design instructions, not measured productivity results.

MarkTechPost’s report linked the prototype and its documentation on November 1.