DeepSeek open-sources dsh-superagent, a coding agent that takes work from GitHub and team chat

The project's stated rule is to "take as little human attention as possible, and give as much value as possible"[2]. In practice the agent confirms a request once, stays quiet while it works, and writes again only when a result is ready, it needs a decision, or it can't finish[2]. It asks one concrete question at a time and is told to mark drafts as unfinished and report failures as failures[2].

According to the South China Morning Post, the design came from a technical article by DeepSeek's Harness team and aims to cut the human work of coordinating several AI agents and checking on their progress[4]. The Next Web, citing Bloomberg, reports DeepSeek is pointing it at Anthropic's Claude Code[5].

Work starts only when a person on an allowed list addresses the agent, or when someone writes in a conversation the team opened to requests[2]. The host enforces those lists, rules and per-task mandates, and the docs say messages, comments and quoted documents are only inputs to a decision and can't change those permissions[3]. Bots can at most produce an "inferred" candidate, never an explicit request[3]. Work nobody asked for needs a rule the team wrote and starts in a mode that only records what the agent would have done[2].

DeepSeek says much of the engineering was done by coding agents, the code may contain bugs and inconsistent style, and the public version differs from its internal one; some designs in its article are future directions not built into the prototype[1]. Setup is long enough that the README tells you to have your own coding agent walk through SETUP.md with you[1]. The operator guidance is blunt too: start with a short allowed list, keep proactive work in record-only mode until it looks right, and don't use the agent to churn out mass pull requests[2].

Most coding agents still live in a terminal or IDE and wait for one developer. dsh-superagent moves the agent into the team's existing threads and treats human attention as the scarce resource, with permissions enforced by the host rather than by whatever text lands in a chat. That's a pattern worth studying for anyone running agents on a shared repo, even if this release is explicitly a prototype to read and adapt rather than deploy as-is.