DeepSeek details a Super-agent architecture and open-sources a prototype: a resident Router dispatching parallel Workers
On 31 October 2026 the DeepSeek Harness team published “Exploring Super-agent Design,” an architecture for coding agents that break away from a single session and stay online to move several pieces of work forward at once: a Router layer handles people and task assignment, Worker sessions execute in isolation, and a Watchdog monitors progress. The same day it released the dsh-superagent-prototype on its official GitHub under the MIT license, stating that it is research code and not an official DeepSeek product.
According to IT Home’s summary, DeepSeek argues that as models need less human help on individual tasks, the bottleneck shifts to human attention: today’s agents are organized around single sessions whose context users must switch between and maintain by hand, and the team observes that a person can usually follow only three to five active sessions at once. The article defines four traits of a Super-agent: continuity (one agent across chat, GitHub comments and other channels), efficiency (several complex tasks in parallel), proactivity (watching environment events and acting when worthwhile) and persistence (always online and able to recover from failures).
A resident Router layer takes all human interaction, keeps global context and assigns tasks, while Worker sessions run concurrently in isolated contexts and report back through the Router. External events such as chat messages, email, code-review comments and CI results are filtered first by program rules and then by a model deciding whether to act; once a Worker owns a pull request it can subscribe to follow-up comments directly instead of going through the Router each time. The article also describes speculative execution (reproducing a problem and looking for a fix in the background, staying quiet if nothing useful turns up), an asynchronous ReAct loop that DeepSeek says is implemented in DeepSeek Harness but not enabled by default, and separation of compute, session state and execution environment: state lives in the cloud and the user’s machine runs only a lightweight daemon, so work survives the device going offline. A separate Watchdog session checks whether each Worker is making real progress or drifting from its goal.
In DeepSeek’s preliminary experiment, a simulated environment with about 300 messages and about 20 work requests was run five times per setup: the Super-agent’s median task completion time was 12.4 minutes and its 95th percentile 15.7 minutes, against a 95th percentile of 29.5 minutes for a single session with subagents under a neutral prompt. QbitAI reports that removing event pass-through raised Router decisions by about 30% and cache-read tokens by about 37%. DeepSeek notes the tests use synthetic load in a simplified environment and measure scheduling and coordination, not the ability to execute individual tasks; the article lists open problems including scaling the Router, long-term memory, cross-device permissions and prompt injection.
The released dsh-superagent-prototype is a proactive coding agent for one GitHub repository, packaged as a DeepSeek Harness bundle. It listens to GitHub (and optionally Feishu); when an allowed person asks for work it opens a work item, does it in its own session and directory, and answers in the original conversation, for example with a pull request, while a web panel shows what needs a person. The README stresses that it is a research prototype for reference only, built largely by coding agents and possibly buggy; the published version differs from the internal one and from the article’s implementation, some designs in the article are not implemented, and users should review it before use.
Sources
- deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent System — DeepSeek (GitHub)
- DeepSeek 提出 Super-agent 架构设计：Router 与 Worker 双层架构，探索更自主的 Agent 系统 — IT之家
- 刚刚，DeepSeek重磅开源：Agent底座架构！ — 量子位