DSH Superagent

Status check 2026-11-13 — active

Verdict: Under the 30-day rule it counts as active: the only commit landed on 2026-10-31, 13 days ago, and GitHub shows 22.9k stars and 2.8k forks. Last release: no releases (commits only); the repo has no tags and no GitHub releases · last commit: 2026-10-31 · licence: MIT (LICENSE file at the repo root, linked from the README; GitHub reports MIT) Repo: https://github.com/deepseek-ai/dsh-superagent-prototype Changed since this note: The note dates from 2026-11-07 and the repo has had no commits since 2026-10-31. One open item from the note is now settled: press coverage (SCMP and Interesting Engineering, 2026-11-01) confirms the companion article ‘Exploring Super-agent Design: Towards a More Autonomous Agent System’.

🗒️ deepseek-ai/dsh-superagent-prototype — a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu. When an allowed person asks for work, it opens a work item, handles it in its own agent session and directory, and replies in the original conversation, for example with a pull request. A web panel shows what needs a person. MIT.

⚠️ Research prototype, reference only, not an official DeepSeek product. Much of it was written by coding agents, may contain bugs, and differs from DeepSeek’s internal version and from the design article Exploring Super-agent Design: Towards a More Autonomous Agent System. Some designs in the article are future directions and are not implemented here.

Description

🧩 What it is:

  • Philosophy: as little human attention as possible, as much value as possible. An agent that can act for a whole team can also waste a whole team’s attention, so it acts only when asked or when a rule the team wrote allows it, and speaks only when someone must act or can use a result.
  • Work items — each request becomes one item with its own agent session and working directory; results come back in the conversation it started from.
  • Authorization — only allowed people can ask for work (doc 002).
  • Sources — GitHub, optionally Feishu; the cookbook explains adding another Source.
  • Panel — a web view of what needs a human decision.
  • This is the original that Router and Workers — DeepSeek Super Agent in Claude Code (AI LABS) rebuilds in Claude Code by copying the router/worker ideas. The AI LABS setup is not this code, and the README does not use the router/worker/benchmark wording from that video, so those claims still need the design article.

Download or use

  • Setup is long, so an agent does it with you. Tell your coding agent: “Read SETUP.md in this repository and set up dsh-superagent with me.” It asks one short question at a time, shows the whole plan once, then installs and configures.
  • Docs, read in order: 000 Philosophy, 001 Domain model, 002 Authorization, 003 Execution, 004 Prompts, 005 Panel, 006 Deployment, 007 Architecture, 008 Cookbook.
  • Dev check: pnpm check (typecheck, tests, build).

Reasoning for

  • Reads like the reference design behind the “one agent you talk to” idea: compare my fleet notes Orca, Swarm Research — Orchestrating Coding Agents and the ticket-first approach in Specification-Driven Development.
  • Worth reading the philosophy and authorization docs even without running it: the attention-budget rule is a useful test for any agent I build (Harness Engineering).
  • Not for me as-is: it targets GitHub (plus Feishu) and DeepSeek Harness, not my Claude Code setup.
  • todo find and read the design article Exploring Super-agent Design and check the router/worker claims from the AI LABS video against it.

Alternatives considered

  • Orca — worktree-per-agent desktop fleet manager, human-driven rather than proactive.
  • OpenRig — persistent multi-agent teams in YAML.
  • The router/worker pattern rebuilt by hand in Claude Code.

Resources


Template: tool