DeepSeek Open-Sources the dsh-superagent Multi-Agent Architecture
DeepSeek has open-sourced the dsh-superagent multi-agent architecture, turning code repositories into automated pipelines: each repository is assigned a resident overall dispatcher (Router Session), which only reviews GitHub Issues, PRs, or CI errors and does not directly modify business code; after a task passes whitelist evaluation, the system spins up a separate Worker for it, assigning an independent session, an independent directory, and a git worktree branch for parallel execution. The system opens 3 work slots by default, supports a maximum of 1 to 16 concurrent tasks, and always reserves 1 slot for humans to jump the queue; subsequent events for the same task are subscribed directly to the corresponding workstation, bypassing the overall dispatcher to avoid repeatedly consuming Tokens. The author says it takes just 5 steps to get started.
Why read this
AI score 73
The original article breaks down how the overall dispatcher and parallel workstations divide responsibilities, as well as the design trade-offs of concurrency slots and direct event routing.
Topics
View original postEnglish text is being prepared.
View original
DeepSeek 刚刚开源的新项目dsh-superagen， 它展示了如何用最少的人类注意力， 换取最大密度的工程交付， 把复杂开发任务的耗时 从将近半小时直接压缩到十五分钟， 这标志着 AI 编程彻底告别了对话框时代， 它不再是一个等你下指令的辅助工具， 而是一个自己去 GitHub 领任务、 多智能体并行开工、自己提 PR 的全自动员工。 它不是让你在网页里粘贴报错信息，而是以独立机器人的身份常驻在你的 GitHub 仓库和飞书群里， 只要团队成员在 Issue 里提了一个需求，或者在群聊里讨论了一个 Bug，它就会自动捕捉信号并生成不可篡改的工作工单，自己创建独立的会话环境去写代码、跑测试，最后直接带着通过测试的代码回到原对话里提交拉取请求。 为了解决多任务堵车和上下文混乱的问题，DeepSeek 在底层设计了一套分流与执行的多智能体架构。 主路由只负责给授权任务分派工单，具体的代码修改则会直接分发给多个并行的工作节点， 每个节点通过独立的 Git 工作树进行物理隔离， 实测把复杂任务的完成耗时直接从二十九点五分钟腰斩到了十五点七分钟。 后续所有针对这个拉取请求的代码评审意见，系统都会自动订阅并路由回原本负责的工作节点，全程不需要人类开发者反复手动召唤。 这个项目真正的野心，是把工程师从琐碎的 Issue 沟通和修补流程里彻底解放出来，让 AI 真正成为一个只需要极少人类注意力、却能持续输出工程价值的异步开发团队。
Body · AI translation
Once you understand dsh-superagent, the multi-agent architecture DeepSeek just open-sourced, you alone become a super-individual with an automated outsourced engineering department. It can be called a landmark engineering milestone that completely liberates programmers from being beasts of burden staring at sessions, and you can get started in 5 steps.
It is not yet another code assistant that chats with you one-on-one in a web chat box, nor is it a mindless script that constantly falls into infinite loops on its own.
The design goal of dsh-superagent is extremely cold and can be stated in one sentence: minimize human attention as much as possible, and squeeze out the greatest possible code value.
In the past, when using AI to write code, the biggest pain point was acting as a human supervisor:
keeping the session window open to watch progress, and you had to review every segment it wrote,
if you wanted it to fix two bugs at the same time, the entire terminal would directly jam,
and once an error occurred or it went off track midway, a person had to frantically clean up the mess behind it, which consumed more effort than writing the code by hand.
DeepSeek's architecture directly turns the code repository into a fully automated assembly line. Each repository is equipped with one resident chief dispatcher (Router Session), which never personally touches specific business code, and is only responsible for sitting behind the gate and reviewing GitHub Issues, PRs, or CI errors;
once a task passes the whitelist gate evaluation, the system immediately opens a separate workstation (Worker) for that task, assigning it a dedicated session, an independent directory, and a git worktree branch, with multiple tasks starting work fully in parallel without disturbing one another.
The most ruthless part is that it tightly controls resources and channels: by default the system opens 3 work slots and supports at most 1 to 16 concurrent tasks, but the bottom line always strictly reserves 1 dedicated slot for humans to cut in and send messages;
new events subsequently generated by the same task go directly through subscription bypass, skipping the chief dispatcher and being sent to the corresponding workstation, never wasting extra Tokens on repeated decisions;
as long as no human decision is needed, it silently switches branches, finishes running tests, and submits PRs in the background, calling you only when it is stuck or when delivering.