DeepSeek 开源 dsh-superagent：采用 Router-Worker 架构的代码智能体原型

DeepSeek Open-Sources dsh-superagent: A Router-Worker Coding Agent

文章背景与核心概要

在当前的软件工程与自动化浪潮中，主流代码智能体多依赖单会话模式工作，面对高并发的消息轰炸与复杂的任务流时，开发者的注意力往往成为制约效率的最大瓶颈。DeepSeek 团队正式开源了研究级原型项目 dsh-superagent，创新性地采用“路由调度-多工并发 (Router-Worker)”架构。主路由长驻监听来自 GitHub 与 Feishu 的事件流，并按需解耦为完全隔离的独立工单 (Tickets)，交由各执行节点并行推进，不仅能主动感知未指派的潜在需求，还能全自动化静默执行并交付拉取请求 (Pull Request, PR)。模拟测试显示，该框架在海量并发请求下将任务交付时长缩短了近一半，展示了未来全自主代码仓库管家的全新范式。

📋 概要

📋 Summary

DeepSeek Harness 团队正式开源了 dsh-superagent。这是一个基于 MIT 协议发布的研究原型项目，专为单个 GitHub 代码仓库打造全自动、主动式编程 AI 智能体 (AI Agent)。该系统采用了创新的 Router-Worker 架构，能够实时监听来自 GitHub 与 Feishu 的事件消息流，并智能地将复杂需求拆解为相互隔离、支持并行处理的独立任务工单 (Tickets)。仿真测试数据显示，在高负载的消息洪峰下，这种超级智能体架构的任务交付速度比传统的单会话基线方案快了将近一倍 (p95 耗时仅为 15.7 分钟，而基线方案需要 29.5 分钟)，有效化解了 AI 辅助编程场景中人类工程师精力无法兼顾多任务的注意力瓶颈。

The DeepSeek Harness team has open-sourced dsh-superagent, an MIT-licensed research prototype designed to act as a proactive coding agent for a single GitHub repository. Operating on a Router-Worker architecture, the agent listens to GitHub and Feishu event streams, automatically breaking down requests into parallelized, isolated work items (Tickets). Simulations indicate that the super-agent framework completes tasks nearly twice as fast as single-session baselines (p95 of 15.7 minutes vs. 29.5 minutes) under heavy message streams, mitigating human attention bottlenecks in AI-assisted development.

📌 核心速览 (TL;DR)

📌 TL;DaR (Too Long; Didn't Read)

• 它是什么？ 一套基于 MIT 开源协议发布的 TypeScript v0.1.0 代码包，可灵活适配 DeepSeek Harness 配置支持的任意底层大模型。
• 环境要求：以专用的 DeepSeek Harness 配置文件形式在本地运行。系统需预装 Node.js (22.19+ 或 24+)、Git (2.32+)、Python 3.8+、tmux 以及 GitHub CLI (2.48+)。
• 性能表现：在处理高并发、杂乱的消息流时，效率显著超越传统的单会话智能体。
• 最佳战绩：任务完成耗时 p95 为 15.7 分钟，对比单会话基线的 29.5 分钟，提速约 1.9 倍。
• 当前局限：目前主要在合成消息流和模拟工具环境下完成测试，在真实工业级编程场景中的代码交付质量尚待进一步检验。

• What is it? A TypeScript v0.1.0 bundle released under the MIT license, adaptable to any model supported by DeepSeek Harness profiles.
• System Requirements: Runs locally as a dedicated DeepSeek Harness profile. Requires Node.js (22.19+ or 24+), Git (2.32+), Python 3.8+, tmux, and GitHub CLI (2.48+).
• Performance: Handles noisy message streams significantly faster than single-session agents.
• Best Case: p95 task completion in 15.7 minutes vs. 29.5 minutes for the single-session baseline (~1.9x faster).
• Limitations: Tested primarily on synthetic messages and simulated tools; real-world coding quality remains unmeasured.

什么是 DeepSeek dsh-superagent？

What is DeepSeek dsh-superagent?

dsh-superagent 是一款既能“有求必应”又兼具“主动出击”能力的 GitHub 智能体。一旦被授权用户唤醒，它便会立即创建一个名为 Ticket (任务工单) 的专属工作单元。每个工单都会在完全独立、隔离的会话和工作目录中运行，互不干扰，最终将执行成果——通常是一份准备就绪的拉取请求 (Pull Request, PR) ——自动汇报给最初的发起会话。此外，它还配备了一个本地 Web 操作看板，专为少数确实需要人类介入决策的节点提供集中管控入口。

dsh-superagent is a reactive and proactive GitHub agent. When triggered by an authorized user, it initializes a dedicated work item called a Ticket. Each Ticket executes in its own isolated session and directory, returning results—often in the form of pull requests—back to the original conversation thread. A local web panel provides a centralized interface for tasks requiring human intervention.

根据官方 README 与 设计哲学文档 (Philosophy Documentation) 的说明，系统的核心工程哲学在于：在创造最大实用价值的同时，将人类的注意力开销降至最低。智能体在接收并确认需求后便会“静默深耕”，在代码编写和调试过程中绝不无故打扰用户；唯有在最终呈现成果、遇到必须由人裁决的阻塞卡点，或是汇报执行异常时，才会主动与开发者交互。

According to the official README and Philosophy Documentation, the core design philosophy centers on maximizing value while minimizing human attention overhead. The agent confirms requests once, operates silently during execution, and communicates only when presenting results, blocking questions, or reporting failures.

DeepSeek 为何要打造超级智能体？

Why Did DeepSeek Build a Super-Agent?

当前主流的 AI 编程智能体通常把“单一交互会话”作为不可分割的工作基本单元。然而在实际工作场景中，人类程序员能够同时兼顾、有效跟踪的活跃会话上限通常只有 3 到 5 个。正如相关研究与 凤凰网报道 (reported by ifeng) 所强调的那样：在多任务并发的智能体工作流中，制约整体效率的头号瓶颈已不再是底层大模型的能力上限，而是人类极其有限的注意力资源。

Mainstream AI coding agents traditionally treat a single session as the atomic unit of work. However, human developers can effectively monitor and manage only 3 to 5 active sessions simultaneously. As highlighted in research reported by ifeng, human attention—rather than model capability—becomes the primary bottleneck for multitasking in agentic workflows.

Router-Worker 架构是如何运作的？

How Does the Router-Worker Architecture Work?

该研究原型实现了一种优雅解耦的双层分工架构：

The research prototype implements a decoupled, two-layer architecture:

1. 路由器 (Router)：负责全天候长驻运行，专门接管外界交互、研判事件价值并进行工单派发；
2. 工作节点 (Workers)：各个相互隔离的执行会话，专职并行执行被指派的具体编码任务。

1. The Router: A long-running session responsible for handling interaction, event evaluation, and task dispatching.
2. Workers: Isolated sessions that execute specific tasks in parallel.

根据 领域模型文档 (Domain Model Documentation) 的定义，整体工作流转过程如下：

As defined in the domain model documentation, the workflow proceeds as follows:

• 数据源 → 触发器 (Source → Trigger)：系统自动摄入并监听来自 GitHub 或 Feishu 的事件消息；
• 评估机制 (Evaluation)：系统审视并研判传入的触发事件是否值得转化为新的开发工作；
• 工单创建 (Ticket)：一旦确认立项，系统便生成一个获得明确授权的工作单元，并附带一份不可篡改的 任务授权书 (Mandate)，明确划定其职责范围与权限边界；
• 事件订阅 (Subscription)：后续若在相关拉取请求 (PR) 下追加讨论或评论，系统会自动将其路由至负责该任务的原专属 Worker 会话，无需人类开发者反复重新 @ 智能体。

• Source → Trigger: Events are ingested from GitHub or Feishu.
• Evaluation: The system assesses whether incoming triggers warrant new work.
• Ticket: An authorized unit of work accompanied by an immutable Mandate outlining its scope and permissions.
• Subscription: Follow-up comments on pull requests are automatically routed to the correct Worker without requiring fresh mentions.

在具体运行中，主会话加载轻量级工具集的 dsh-superagent-router 预设，而各个具体的 Worker 节点则执行 dsh-superagent-worker。正如 执行指南 (Execution Guide) 中所强调的，为了保证主交互通道永远保持丝滑顺畅与持续响应，所有 Shell 命令行操作在执行满 5 秒后都会被自动转入后台异步运行。

The main session runs the dsh-superagent-router preset with a lightweight toolset, while individual Workers execute dsh-superagent-worker. To maintain continuous responsiveness, shell commands are automatically backgrounded after 5 seconds, as outlined in the execution guide.

智能体的主动执行行为受到极其严格的用户权限约束。根据 权限授权文档 (Authorization Documentation) 的规定，系统提供了三种接收模式：

Proactive operations are strictly governed by user permissions. The authorization documentation specifies three accept modes:

• off：完全关闭主动操作；
• observe (默认模式，智能体只在日志中记录其“本打算”采取的操作，而不实际执行)；
• auto：完全开启自动执行模式。

• off
• observe (default mode, which logs actions the agent would have taken)
• auto

性能测试与效果评估

Performance & Evaluation

DeepSeek 团队利用模拟群聊和 GitHub CI 事件流对该系统进行了严格评测。在连续 5 轮迭代测试中，每轮测试均注入了约 300 条模拟消息 (其中穿插了约 20 条真实的有效开发请求)。

DeepSeek evaluated the system using simulated group chat and GitHub CI event streams, with each run containing approximately 300 messages (including ~20 legitimate work requests) across five iterative test runs.

• 超级智能体表现：任务交付周期的 p50 为 12.4 分钟，p95 为 15.7 分钟；
• 单会话 + 子智能体 (中性提示词基线)：任务交付周期的 p95 为 29.5 分钟 (未披露 p50 数据)。在中性提示词条件下，基线模型在全部 5 轮测试中调用子智能体的次数为 0 次；
• 路由器负载伸缩能力：当外部事件到达速率从每分钟 1 次激增至 40 次时，系统的平均批处理规模从 1.0 增至 2.8，而单次请求的处理延迟仅从 2.5 秒温和微增到 3.9 秒；
• 消融实验 (Ablation Study)：若移除事件直接透传通道，将导致路由器的决策次数增加约 30%，同时缓存读取 Token 数量增加约 37% (从 2323 万 Token 暴增至 3177 万 Token)。

• Super-Agent Performance: p50 of 12.4 minutes; p95 of 15.7 minutes.
• Single Session + Subagents (Neutral Prompt): p95 of 29.5 minutes (p50 metrics were not disclosed). Under neutral prompts, the baseline model invoked subagents 0 times across 5 runs.
• Router Load Scaling: As arrival rates increased from 1 to 40 events per minute, average batch sizes grew from 1.0 to 2.8, while per-request latency increased modestly from 2.5 to 3.9 seconds.
• Ablation Study: Removing direct event pass-through resulted in a ~30% increase in router decisions and a ~37% increase in cache-read tokens (jumping from 23.23M to 31.77M).

• (注：以上所有数据均为基于该研究成果二次分析所得的自测公开数据。) *

(Note: All figures are self-reported via secondary analyses of the research findings.)

横向对比一览：dsh-superagent 与其他编程智能体

Comparative Overview: dsh-superagent vs. Other Coding Agents

核心要点总结

Key Takeaways

• 开创性的主动式原型：DeepSeek 发布的 dsh-superagent 是一份基于 MIT 协议的开源蓝图，展示了实现自主化代码仓库治理的前沿探索；
• 解耦式并发扩展：通过核心中央路由器进行任务智能研判与派发，并在完全隔离的 Worker 节点中并行执行，实现了优雅的横向扩容；
• 显著的效率跃升：在模拟高负载测试中，任务交付效率大幅领先传统模式 (p95 耗时仅 15.7 分钟，而传统单会话基线高达 29.5 分钟)；
• 安全防护与沙箱隔离：由于智能体在执行时拥有完整的用户级操作系统权限，在实际生产环境中部署前，强烈建议配置严格的安全隔离与沙箱运行机制。

• Proactive Prototype: DeepSeek released dsh-superagent as an MIT-licensed blueprint for autonomous repository management.
• Decoupled Scaling: Work is intelligently dispatched via a central Router and executed concurrently across isolated Worker sessions.
• Efficiency Gains: Simulated workloads demonstrated significantly faster task completion (p95: 15.7 mins vs. 29.5 mins for traditional single sessions).
• Security & Isolation: Because the agent executes with full user-level OS access, proper containment and sandboxing are strongly recommended before deployment.

相关资源与项目链接

Resources & Links

• 📂 GitHub 开源代码库：deepseek-ai/dsh-superagent-prototype
• ⚙️ 快速安装与配置指南：SETUP.md
• 📄 系统架构与设计文档：docs/

• 📂 GitHub Repository: deepseek-ai/dsh-superagent-prototype
• ⚙️ Setup Guide: SETUP.md
• 📄 Design Documentation: docs/

原文发表于 MarkTechPost。

Original post published on MarkTechPost.