DeepSeek’s Super-agent Prototype Ships, but Watchdog Remains Unimplemented
DeepSeek’s Harness team published a Super-agent architecture article and open-sourced the TypeScript dsh-superagent-prototype under the MIT License on October 31, 2026. A Juejin analysis of the repository found that the prototype implements an admission table, immutable Mandate, shadow mode, two-stage event filtering, and revalidation before recovery, but not the proposed external Watchdog or storage-compute separation. The architecture article’s performance results used synthetic workloads and a simplified environment, so the team presented them only as preliminary references.
The repository provides a concrete implementation of authorization controls for agentic systems while clearly showing that the proposed Watchdog remains future work.
2026-10-31,DeepSeek Harness 团队发布架构文章《Super-agent 设计探索:迈向更自主的 Agent 系统》,同日开源原型 deepseek-ai/dsh-superagent-prototype(MIT / TypeScript)。三天 1.36 万 star。
大多数转述停在了"AI 会自己找活干了"。我做了点不一样的事:把仓库 README 和 docs/000/002/007 逐文件拉下来读(钉死在唯一 commit 6337fc8),再把官方公众号原文逐字核定了一遍。这篇文章讲三个转述里看不到的东西:原型里真正跑起来的授权链长什么样、文章的实验数字应该怎么读、以及一个重要落差——文章的主角 Watchdog 在原型里没有实装。
一、30 秒版本
- 立论:"单任务的上限取决于模型能力,多任务的上限取决于人的注意力"——团队观察到一个人通常只能同时跟进 3 到 5 个活跃 Session;
- 方案:Router/Worker 双层 + 事件驱动(两级过滤)+ 异步多信道 + 存算分离(BYOM)+ 外部 Watchdog,五个模块支撑"连续/高效/主动/持久"四个特征;
#DeepSeek#DeepSeek Harness#dsh-superagent-prototype#Super-agent architecture