DeepSeek Super-Agent is now public code. On Oct. 31 the company’s Harness team described the design on WeChat and published dsh-superagent on GitHub under an MIT licence, calling it a research prototype and “not an official DeepSeek product.” Bloomberg’s Debby Wu framed the release as a challenge to Anthropic’s Claude Code, The Next Web reported.
Key Takeaways
- DeepSeek released dsh-superagent, an MIT-licensed research prototype, not an official product.
- It listens to one GitHub repo and optional Feishu, often replying with a pull request.
- Simulated p95 task time was 15.7 minutes versus 29.5 for one session.
- Default mode is observe; it runs with full OS-user access, so isolate it.
What DeepSeek Super-Agent Watches and Returns
It is a TypeScript bundle, v0.1.0, one commit, packaged for a dedicated DeepSeek Harness profile on your own machine, MarkTechPost reported. It uses whatever model the profile is set to. Scope is one GitHub repository, plus optional Feishu, ByteDance’s workplace chat.
An allowed person asks; the agent opens a Ticket, works it in its own session and directory, and replies in the thread, often with a pull request. A local web panel lists only items that still need a person. The README says some design-article ideas are not built yet, and tells users to review and adapt the code first.
MarkTechPost lists these local requirements:
- Node.js 22.19+ or 24+
- git 2.32+
- Python 3.8+
- tmux
- GitHub CLI 2.48+
Router-Worker Dispatch, Gated by Default
The Harness team argues that mainstream agents treat one session as the unit of work, and that a person can actively follow only three to five sessions, so attention, not model capability, is the bottleneck, ifeng’s account of the report said. Ben Jiang, writing in the South China Morning Post, reported the Harness team’s stated aim to reduce the need for humans to coordinate multiple agents and monitor their progress.
- A long-running Router (dsh-superagent-router, smaller tool set) handles interaction and dispatch.
- Isolated Workers (dsh-superagent-worker) execute each Ticket in its own session.
GitHub or Feishu events become Triggers. Eval decides if a batch asks for new work. A Ticket holds an immutable Mandate stating why work may proceed.
Subscriptions route later pull-request comments to the right Worker without a fresh mention. Shell commands background after five seconds so Workers still receive messages.
Bloomberg’s account says it can suggest solutions without being summoned. Accept modes are off, observe (default) and auto. In observe, it only records what it would have accepted.
The philosophy document says it acts only when asked, or when a team’s own rule allows it. The GitHub manual states the rule:
Take as little human attention as possible, and give as much value as possible.
It confirms once, stays quiet while working, and writes again only with a result, a question or a failure.
Simulated Speed Under Noisy Chat
DeepSeek’s test, relayed by ifeng and summarized by MarkTechPost, used synthetic chat and GitHub CI, with simulated tools. Real coding quality was not measured. Each scheme ran five times, on about 300 messages, about 20 of them real requests.
| Simulated measure | Super-agent | Single session plus subagents (neutral prompt) |
|---|---|---|
| p50 task time | 12.4 minutes | No p50 given |
| p95 task time | 15.7 minutes | 29.5 minutes |
| Runs | 5 | 5 |
The single-session baseline, under a neutral prompt, called a subagent zero times in five runs. MarkTechPost links that to the slower p95.
Router load rose with traffic: from about one to 40 arrivals a minute, batch size went from 1.0 to 2.8 and per-request latency from 2.5 to 3.9 seconds. Dropping direct event pass-through lifted Router decisions about 30 percent and cache-read tokens about 37 percent, from 23.23 million to 31.77 million. All of it is self-reported.
Treat the roughly 1.9x p95 gap as a scheduling result on synthetic noise, not a quality score: the baseline never fanned work out.
Claude Code, Funding, and US Limits
The Straits Times, citing Bloomberg, said DeepSeek built the Harness team earlier in 2026 to take on coding services from US labs, including Anthropic. In August DeepSeek released a preview of a Claude Code rival, an open-source app now in public preview for coding, documents and spreadsheets, The Next Web reported.
At GitHub Universe, GitHub showed a Copilot app running agent work across several repositories in parallel sessions, a wider surface than this one-repo prototype. See also our report on GitHub Copilot Fusion going generally available. CooperBench, a January study that did not test DeepSeek’s agent, found team success rates on average 30 percent lower than when each agent worked alone.
Ahead of a planned listing, The Straits Times, citing Bloomberg, put the latest raise at about 100 billion yuan (S$19.2 billion). The Next Web, also citing Bloomberg, said about $15 billion. Civilian US federal agencies were told last month to remove DeepSeek.
MarkTechPost notes the prototype runs with full OS-user access, so isolate the profile. For the ablation’s token jump, see our report on TokenEconomizer and coding-agent costs.
