DeepSeek dropped dsh-superagent on October 31st: an open-source, MIT-licensed coding agent that parks itself in your GitHub repository, monitors incoming issues, and submits pull requests without being explicitly prompted. Headlines calling it a “Claude Code killer” are wrong — but the Router-Worker architecture underneath is worth understanding before the rest of the industry copies it.
DeepSeek’s New Agent Doesn’t Wait to Be Asked #
Most AI coding agents are interactive: you open a terminal, type a prompt, and wait. Super-Agent breaks that pattern. It runs as a persistent background service. When an authorized user opens a GitHub issue, the agent picks it up autonomously, spins up an isolated worker session, executes changes in a dedicated directory, and replies in the same thread with a pull request. No one summoned it.
Optionally, it also listens to Feishu — ByteDance’s workplace chat tool — handling requests from the same conversation where your team is already discussing the problem. The result goes back to the same thread.
“Take as little human attention as possible, and give as much value as possible.”
DeepSeek Super-Agent README
The Router-Worker Architecture #
The real contribution of DeepSeek Super-Agent is the design pattern, not the model. DeepSeek built a two-layer system for parallel, autonomous task execution:
- Router: Lightweight, small tool set. Monitors GitHub and Feishu for incoming events. Uses two-level event filtering — program rules screen first, only ambiguous events go to a model for judgment.
- Workers: Full tool permissions, each running in an isolated directory. Each task gets its own session. Workers lose their tools when the ticket ends — no persistent state between tasks.
- Watchdog: An AI system that supervises workers, detecting failures, infinite loops, and goal drift before they compound.
The throughput numbers are credible. At roughly 40 triggers per minute, latency increases from 2.5s to just 3.9s. Per-decision inference time drops from 6.4s to 1.6s as the Router batches work. These aren’t benchmark claims against competitors — they’re internal design metrics, and they’re the kind of detail that separates a prototype from a toy.
The Attention Bottleneck #
Developers can manage roughly three to five active agent sessions at once. That’s a human constraint, not a software one. Super-Agent doesn’t eliminate human oversight — it shifts what you review. Instead of context-switching between terminal windows and prompt boxes, you’re reviewing completed pull requests in GitHub. The bottleneck moves from generating code to approving it.
That reframe is subtle but real, and it’s the reason background agents like Super-Agent occupy a different niche than interactive tools like Claude Code or Cursor. Both have a place in a mature agent stack.
Security: Read This Before You Run It #
DeepSeek is explicit: this is a research prototype. The caveats are not boilerplate.
- The blocked-command list is a default, not a hard security boundary. Workers can run programs by absolute path.
- Workers execute as the operator’s OS user. DeepSeek recommends a dedicated OS account before connecting the agent to any real repository.
- GitHub issues can contain adversarial content. Prompt injection via issue bodies is a genuine attack surface — any user who can open an issue can attempt to manipulate the agent.
- DeepSeek listed four unresolved open problems themselves: scaling the Router, collaboration UX without a chat window, context boundaries at org scale, and securing cloud agents with local machine access.
Test this in a sandboxed environment. Do not point it at production repositories until the security model matures. See the official deployment guide for the recommended isolation setup.
How to Run DeepSeek Super-Agent #
Requirements: Node.js 22.19+, GitHub CLI 2.48+, tmux, pnpm, Python 3.8+, and DeepSeek Harness already configured with a model provider.
npm install -g @deepseek-ai/dsh@0.2.1-rc.3
git clone <repo> dsh-superagent && cd dsh-superagent
pnpm install && pnpm build
scripts/dsh-superagent start
You’ll need to configure cordis.patch.yml with your repository paths, allowed GitHub accounts, and execution settings. The setup is designed to be done alongside a coding agent — the deployment guide walks through the full process step by step. Feishu integration is optional and requires lark-cli.
Worth Watching, Not Yet Worth Depending On #
Super-Agent is genuinely interesting and genuinely unfinished. The Router-Worker pattern is solid architectural thinking for background agent systems. The GitHub-native PR workflow is more practical than most “autonomous agent” pitches from the past two years.
But there are no published benchmarks against Claude Code, Devin, or OpenHands — and DeepSeek’s own disclaimer that the release is “not identical to our internal version” is an honest signal about the gap between prototype and product. It accumulated 3,400+ GitHub stars within 24 hours, which reflects developer appetite for background agents more than production confidence.
Give it 30 days. If you’re evaluating async coding agents now, clone it into a sandbox repo and run it on low-stakes issues. That’s the right level of commitment for v0.1.0.