DeepSeek Open-Sources dsh-superagent: the Super-agent Blueprint Goes Public

The facts, as verified (Oct 31)

• The repo, verified directly. deepseek-ai/dsh-superagent-prototype, description "Towards a More Autonomous Agent System", MIT license, created 2026-10-31T09:32:46Z, 1,500+ stars within the day. The README's first line frames everything: research prototype, reference only, not an official DeepSeek product, grown out of the team's day-to-day development with much of the engineering done by coding agents.
• What actually runs. A proactive coding agent scoped to one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu. When someone allowed asks for work, it opens a work item, works on it in its own agent session and directory, and answers in the original conversation — for example with a pull request. A web panel surfaces what needs a person.
• The philosophy, verbatim. "Take as little human attention as possible, and give as much value as possible." The agent 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 — because an agent that can act for a whole team can also waste a whole team's attention.
• The architecture in the writeup (as reported by QbitAI): a Super-agent with four properties — continuous (context across sessions and platforms), efficient (parallel workers), proactive (speculative execution: finding bugs, reproducing and filing fix PRs unprompted) and persistent (24/7 online with self-recovery). The shape is a Router (receives messages, assigns tasks, holds global context) dispatching to a pool of Worker sessions, with compute, state and execution environment separated so the cloud keeps running when a device goes offline. A Watchdog — AI supervising AI — detects failures, useless loops and goal drift, and intervenes.
• Repo internals confirm the split. The main session runs dsh-superagent-router with a smaller tool set; worker subagents run DeepSeek Harness's standard preset with the worker's permissions and lose their tools when the ticket ends. Workers reach GitHub through the agent account's gh and Feishu through lark-cli; a WorkerSlots mechanism reserves capacity before waking idle workers.
• The numbers, and their weight. In the writeup's simulated environment — roughly 300 messages, 20 work requests, five runs per configuration — the Super-agent completed tasks with a median of 12.4 minutes and a 95th percentile of 15.7, against 29.5 p95 for a traditional single-session baseline. An ablation adds the cost angle: removing direct event pass-through raised Router decisions by about 30% and cache-read tokens by about 37%. These are DeepSeek's own measurements of its own design, not an independent benchmark.

Verified against the GitHub repository (read directly on Oct 31, 2026) and the QbitAI report of DeepSeek's engineering writeup. No deployment or benchmark reproduction was performed for this article; completion-time figures are attributed to their source.

Context: the runtime layer is where DeepSeek is investing

Two of the writeup's mechanisms matter as much as the benchmarks. Subscription handoff: once a Worker takes over a thread (say, a GitHub PR), it subscribes to subsequent comments directly instead of routing every event through the Router — the ablation above shows this is a cost feature, not just a latency one. Two-level event filtering: program rules screen events first, and only the ambiguous remainder goes to a model for judgment. Both are token-cost engineering at the architecture layer — the same lever as cached-input pricing, applied in code.

The writeup also describes an async ReAct loop that keeps accepting new instructions while a long task runs, notes it is already implemented in DeepSeek Harness (not enabled by default), and the repo lands exactly there: dsh-superagent is a Harness bundle. DeepSeek keeps investing above the API — Harness as the agent runtime (our multimodal-input coverage), third-party frameworks wiring DeepSeek in (the Hermes-on-DeepSeek stack), and now an opinionated full-stack reference for always-on agents.

On the rumored "dsh dot": QbitAI's framing is that the community got the architecture and a prototype first, with no consumer product announced. Treat that as reported, and treat this release as the signal that actually matters for builders — the design decisions are public, executable and MIT-licensed.

What this means for builders

To try the prototype itself: the repo's setup is deliberately agent-assisted — SETUP.md instructs you to have your coding agent install and configure it with you, one short question at a time. The numbered docs (000 philosophy through 008 cookbook) are meant to be read in order, pnpm check runs typecheck, tests and build, and the MIT license allows full adaptation. Budget for the README's warning: research code, possible bugs, inconsistent style — review before running it against anything you care about.

The transferable part is the pattern set, and none of it requires the prototype: (1) a small-context Router that dispatches, with fat Worker sessions doing the work — context discipline is cost discipline; (2) rules-then-LLM event filtering so cheap code pays for the obvious; (3) direct subscription after handoff so the dispatcher stops paying for threads it no longer owns; (4) an async loop so a 15-minute task doesn't deafen the agent; (5) a watchdog so a 24/7 agent fails loudly instead of silently. All five run on plain OpenAI-compatible DeepSeek endpoints — the same model families covered in the DeepSeek API guide and the comparison table.

Watch three signals for the follow-up: a public URL for the writeup itself, movement toward the rumored consumer agent, and whether Harness bundles become a distribution channel other teams ship into. Any of them moves, this page gets an update entry.

Access from outside China — and the usual disclaimer

This site tracks Chinese model APIs for international builders: official endpoints, pricing and open-source status. The DeepSeek API guide covers the OpenAI-compatible endpoint and model lineup behind this ecosystem; the comparison table holds the cross-vendor ladder.

Independence disclaimer: ChinaModelAPI is an independent information site, not affiliated with DeepSeek or any model vendor. Facts here are drawn from the official repository (read directly) and explicitly attributed reporting; benchmark figures are the vendor's own and unverified; nothing on this page is a routing, availability or product promise.

Primary sources

• GitHub — deepseek-ai/dsh-superagent-prototype (MIT, created 2026-10-31; README, docs 000–008, SETUP.md)
• QbitAI (量子位) report, Oct 31, 2026 — the engineering writeup "Exploring Super-agent Design: Towards a More Autonomous Agent System" and its benchmark details
• 站内对照：DeepSeek Harness 多模态输入 · Hermes on DeepSeek 成本栈 · V4.1-Flash-1017 更新 · DeepSeek API 指南

FAQ (2026)