让AI干活，有一种累，跟自己干还不太一样。
任务交出去了，你还得记着哪个窗口在改代码，哪个窗口等反馈，群里刚补的要求有没有告诉它。隔一会儿切回去看看，做完没有，还是已经停了？
说好的一个人用AI顶一支队伍，结果先给自己升了个职，监工Pro Max。
前两天，DeepSeek产品部发了一篇Super-agent设计文章，盯上的就是这件事：AI越来越会做任务，可好几件事一起推进时，人仍得忙着协调。
它想让AI跟住同一件事，在群聊和代码平台之间接着干，需要你做决定时，再来找你。这次公布的是设计探索和研究原型，不是新模型，也不是打开DeepSeek就能用的新按钮。
团队已经用内部版本跑出了一些例子，咱先看它到底接住了什么。
人换了，地方换了，事情没断
第一个例子是改字体，有人在群里提需求，希望正文、代码和终端的字体能分开设置。维护者叫来Agent，它接下任务，提交了修改。
这里的PR，就是把改好的代码交上去，供其他人检查、讨论，确认后再合进项目。
用户接着问，字号能不能也分开调？Agent把新要求加进了同一个PR。大家转到GitHub审查代码，维护者觉得设置项太多，希望折叠起来。设计师看完又补充，不要另开一块区域，直接从字号设置那里展开。
它一路接着改，最后回到群里，通知大家PR已经合并。
这里省的不是敲几行代码，而是来回搬话。群里多一句要求，评审里换一个方案，要有人告诉AI，大家说的还是同一件事，别重新开工。
本来想让AI帮忙，最后忙着帮AI对消息，谁是谁的助手，还得重新捋捋。这个案例里，人照样提需求、看结果，但不用再兼职传话。
更进一步，没人叫它，它也能先去查问题。
官方展示的另一个例子里，用户反馈macOS窗口没获得焦点时，点一下会话，有时会被误判成双击，突然进入重命名。用户说自己每天能误触几十次，明明只想切个会话，电脑非要给人安排一场起名大会。
这次没人专门@Agent，它跟进反馈，查到了原因：窗口被激活时，第一次点击被底层框架吃掉，接下来那一下却带着双击的计数，界面就误会了。随后它提交修复PR，并在回复里说明附有修改前后的对比。
还有一次，群里在讨论单个波浪号为什么会把文字变成删除线。比如10~20，本来写的是范围，排版却给你另作安排。大家还在聊方案，Agent已经提交了PR，改成单个波浪号保留，两个连续波浪号才表示删除线，并解释处理依据。
这听着像AI学会自己找活了，但它不是见一句话就冲上去。
团队提前给内部实例定了静默复现并修复的规则，它看到相关问题先去尝试，确认问题、找到简单修法，再把结果拿出来，没找到就保持安静。
不然，一个群几十个人，它句句都接，主打一个有问必答，没问也答。问题没少几个，未读消息倒先上去了。
不能只给它加一句：你主动点
想让它接住这些事，先得处理一个很实际的问题，新的要求来了，正在干活的AI听不听得见？
全塞进一个对话，看起来什么都记着，可它正忙着跑任务，新消息又来了，先干哪个？每件事另开一个窗口倒能同时干，但谁知道方案刚改过，谁来统一回复？兜了一圈，又到你了。
DeepSeek的办法，是把接话和干活分开。
在文章展示的双层设计里，前面负责接消息、记住各项任务的关系、分派工作，这一层叫Router，可以理解成调度员。后面负责实际执行的叫Worker，每件任务有自己的执行会话，进度和结果再由Router汇总，对外回复。
你继续交代第二件事，不用等第一件事跑完。
同一件事后续有补充，也不用每次重新挂号。系统记得哪个执行会话在跟这个PR，评审意见就直接送过去，超出当前任务范围的，再交回调度员重新分派。
这里跟几个AI一起讨论一道题不太一样。后者是把一道题做得更好，这里要解决的是，几件不同的工作同时来了，别互相堵着，也别把协调全留给人。
长任务也得插得进话，比如它在等一项检查，你发现要求要改，总不能电脑在跑，人连说话都得排队。
文章里的处理方式是，耗时操作放到后台，先告诉系统已经启动，做完再送回结果。对话不用陪着工具一起等，当前这一步推理结束后，就能接着处理新消息。
这项异步能力在Harness里已实现，但原文发布时还没默认开启。公开原型则把耗时命令转到后台，让Worker能在后续步骤收到新消息。
任务多了，还有一种忙要防，看着一直在运行，其实半天没往前走。设计里因此安排了独立的检查者，既查系统有没有异常，也查任务有没有进展、有没有跑偏。
不然窗口开得再多，最后还得你挨个巡一遍，说好解放双手，先把双眼占满了。
文章还讨论了把会话状态、运行计算和实际干活的电脑分开，让任务不跟着某台电脑的开关机一起消失。上面讲的是整套设计，公开原型只实现了其中一部分，下载代码不等于把全景图里的功能全领走了。
更快了，但省下的时间不是白来的
DeepSeek做了一个初步实验，给系统大约300条群聊和GitHub消息，其中约20条是工作请求，其余混着闲聊、跟进、叫停和检查结果，每种方案重复跑5次。
它测的是怎么安排消息和任务，不是代码写得多好，实验里的编码工作被换成了一个只消耗时间的工具。
消息间隔和工作时长还缩到了原设定的四分之一，模型调用耗时保留真实值，所以下面的分钟数也不是现实项目的交付用时。
先看排在中间的任务，Super-agent用了12.4分钟，对照的单会话方案用了20.4分钟。再看靠后那批慢任务，也就是95分位，两者分别是15.7和29.5分钟。这个对照方案带着子任务工具，用的是中性提示。
等得更短了，后台消耗也更多了，单次实验所有会话合计读取缓存的Token，从对照方案的605万增加到2323万。这里只统计缓存读取，不能直接拿它算账单涨了几倍。
这组实验里还有个挺有意思的细节，给单会话配了分工工具，它却未必真会分工。中性提示下，5次都没调用子任务，全在主会话里自己干，还不断被新消息打断。
分工工具配齐了，干活还是单打独斗。
改成优先委派的提示，情况好了些，可仍有一次觉得事情很小，决定自己来，拖出了长尾。提示词劝它学会分工，它临到干活又觉得，这点小事还用叫人？
DeepSeek这套设计直接把接话和执行分成不同职责，不把是否分工全押在模型当场的决定上。
它在尝试用更多后台计算，换更短、更稳定的等待，以及更少的人类照看。
对一个人来说，值不值就得看，你省下的是偶尔点一下，还是一整天被进度反复打断。
现在能用吗？能试，但别拿主力电脑开盲盒
公开仓库叫dsh-superagent-prototype，目前围绕一个GitHub代码仓库工作，可以接GitHub，也可以接飞书。它更适合已有代码项目、用过编程Agent，也愿意处理部署和权限的人试验，还不是给所有人准备好的通用生活助理。
想试，需要配置好模型的DeepSeek Harness，以及文档要求的开发环境。原型在机器上运行，关机就不能指望它继续干，模型调用和长期运行也得自己承担成本。
官方给的安装入口倒很符合它的思路，把仓库交给正在用的编程Agent，让AI带你装AI。可以直接把下面这段发过去：
# 任务
在GitHub搜索deepseek-ai/dsh-superagent-prototype，阅读该仓库的SETUP.md，按文档带我配置dsh-superagent。
# 执行要求
先核环境、收集必要信息，给出完整计划，等我确认后再安装。
按官方建议从小范围开始，主动任务先用observe，代码由人合并。
需要账号登录或密钥时，给我本地操作步骤，不要让我在聊天里发送密钥。
先定好它在哪个仓库工作、用哪个独立账号、谁能派活，允许名单短一点，代码先由人合并。
主动任务默认先观察，不直接执行。
它会记下哪些消息让自己想接活，你看过这些判断，再决定要不要开放自动执行。
允许名单里的人明确派活，则是另一条路径，观察模式下也能执行。
这道区别很重要，读到消息只是知道了这件事，开始新工作还要有允许的请求，或者命中团队预先写下的规则。否则你随口抱怨一句，它立刻给自己派活，本来想省事，转头变生事。
权限还得看另一层，这个原型执行命令时，拥有运行它的系统账号权限，任务目录和凭据包装不等于安全隔离。
官方文档也建议使用专用系统账号，先用隔离的测试环境、独立账号和不敏感的项目，别把主力电脑上的文件与账号一股脑交过去。
配置完，文档给的最小尝试是，让允许名单里的账号在仓库Issue中@Agent，交一个范围清楚的小任务，去面板看任务是否出现，再检查它交回的修改。
这时候别光看它回复得多热情，改一条要求，看它接不接得住，遇到卡点，看它说不说得清。没叫它的时候，也看它能不能按规矩安静下来。
写在最后
老金觉得，这次值得关注的，不是Agent前面又加了一个Super，而是人的注意力终于被当成了一笔要认真省的成本。
以前切回窗口，是怕它停了。要是以后它来找你，已经把能查的查过、能做的做完，只剩真正需要你拍板的那一步，这一次打断，才算有价值。
所以再看这类工具，除了问它能同时干多少件事，也可以多问一句，这些事交出去以后，我还得回来多少次？
AI忙得过来当然好，别忙到最后，只有人闲不下来。
谢谢你读我的文章。
如果觉得不错，随手点个赞、在看、转发三连吧🙂
如果想第一时间收到推送，也可以给我个星标⭐～谢谢你看我的文章。
开源知识库（实时更新 交流群）：
my.feishu.cn/wiki/OhQ8wq…
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Let's get AI to work. One's tired. It's not like you're doing it yourself.
You'll have to remember which window is changing the code, which window is giving feedback, and whether the demands just made in the group were told. Do you want to go back and see if it's done or stopped?
A man with an AI top team got himself a promotion, supervisor Pro Max.
Two days ago, the DeepSeek Products Department published a Super-Agent design article, which focused on the fact that AI was becoming more and more capable of doing things together, but people were still busy coordinating.
It wants AI to follow the same thing, go between the chat and the code platform, and find you when you decide. This is a design exploration and research prototype, not a new model, or a new button to open DeepSeek.
The team has come up with some examples in the inside. Let's see what it's got.
People change, places change, things stay the same.
The first example is a change in font, where there is a demand in the group to separate text, code and terminal fonts. The maintainer called Agent, and it took the task and submitted the amendments.
The PR here is to hand over the modified code for inspection, discussion, confirmation and integration.
The user went on to ask, could the numbers be split? Agent added the new request to the same PR. You go to GitHub to review the code, and the maintainer feels that there are too many setups and wants to fold them up. When the designer has completed his reading, he adds that he does not open a separate section, which is carried out directly from the font settings.
It went all the way back to the group and informed everyone that PR had been merged.
It's not a few lines of code, it's back and forth. One more line from the group asks that there be another option in the review, and someone should tell AI that it's the same thing to say, not to start work again.
I tried to get AI.'s help. I ended up helping AI. In this case, there is no need for a word of need and results.
Further, no one calls it, and it can start with the questions.
In another example officially presented, when the user feedback MacOS window was not focused, a click on the session was sometimes miscalculated and suddenly renamed. Users say they can miss dozens of times a day, but they just want to have a conversation. Computers have to set up a famous conference.
This time, no one was specializing in @Agent, and it followed up on the feedback and found out why: when the window was activated, for the first time, it was eaten by the bottom frame, and the next time it was double-clicked, the interface got wrong. It then submitted the restoration of the PR and provided a comparison of the changes in its response.
And again, the group was discussing why the individual wave sign turned the text into a striker. For example, 10-20, which was originally written in terms of scope, the layout was arranged for you. You are still talking about the programme, and Agent has submitted a PPR, which has been retained as a single wave, and two successive waves indicate a strikeout and an explanation of the basis for the process.
That sounds like AI learned to find his own life, but it's not just going up in a word.
The team pre-empted the internal case by setting the rules of silent recurrence and repair, and saw the relevant questions first to try, to identify the problem, to find simple modifications, to take out the results and to remain quiet without finding them.
In fact, there are dozens of people whose words will be answered, and the Lord will call and answer without a question. There's a lot of problems. The unreading information goes up first.
You can't just add one word to it: be proactive.
If you want it to take these things, you have to deal with a real problem.
It's all in a conversation, it looks like everything, but it's on a run, new news, which one? Every other window can do it at the same time, but who knows that the plan has just been changed and who's gonna answer it? It's your turn again.
DeepSeek's method was to separate the call and work.
In the two-tiered design presented in the article, the former is responsible for receiving information, remembering the relationship between tasks and assigning tasks, known as Router, which can be understood as dispatchers. The latter, who is responsible for actual execution, is Worker, each mission has its own executive session, and the progress and results are then summarized by Router and responded to externally.
You keep saying the second thing. You don't have to wait until the first thing is done.
The same matter is followed up and does not need to be re-listed each time. The system remembers which executive session is with this PR, and the evaluation is sent directly beyond the current mandate and returned to the dispatcher for reassignment.
It's different here to discuss a subject with a few AIs. The latter is to do a better job, and here it is to solve the problem of several different jobs coming together, not blocking each other, and not leaving coordination to others.
Long missions have to get in line, like waiting for an inspection, and you find that you can't change a computer, you can't run, you have to line up.
The article is handled by putting the operation backstage in a time-consuming manner to inform the system that it has been activated and then returned. Dialogue does not have to wait with the tools, and after this step of reasoning is completed, new information can then be processed.
This staggered capability was achieved in Harness, but the original publish was not opened by default. The public prototypes transfer time-consuming commands backstage so that Worker can receive new information in the next steps.
There's a lot of work to do, and there's a lot of work to do, and it's been running for half a day. The design therefore included independent inspectors to check whether the system was abnormal and whether the mission was moving or not.
Otherwise, you'll have to go around every window, and you'll have to free your hands and fill your eyes.
The article also discussed the separation of the session status, running computing and actual working computers so that the mission would not disappear with a computer switch. It refers to the design package, of which only part was achieved in the open prototype, and the download code does not mean that the functionality in the panorama is removed.
It's faster, but the time saved is not for nothing.
DeepSeek carried out a preliminary experiment to provide the system with approximately 300 sets of talk and GitHub news, of which about 20 were job requests, while the rest were mixed with chats, follow-up, stop and check results, each running five times.
It measured how to organize messages and tasks, not how well the code was written, and the coding in the experiment was replaced by a time-consuming tool.
The message interval and working hours have also been reduced to a quarter of the original set, and the model is used to keep the real value, so the minutes below are not the actual project delivery time.
On the middle line, Super-agent spent 12.4 minutes and 20.4 minutes on a single session programme. The back-up slow assignments, i.e. 95, were 15.7 and 29.5 minutes, respectively. This control program was accompanied by a subtask tool, using neutral tips.
Shorter than that, the backstage consumption was higher, and Token, which was read together by cache, was increased from 6.05 million to 2.23 million for the control programme. Only cache can be read here, and it cannot be counted directly as a multiple increase.
There's also an interesting detail in this set of experiments, with a division of labour tool for single sessions, but it's not necessarily true. On a neutral note, subtask was not called five times, all on its own in the main session, and was repeatedly interrupted by new news.
The division of labour tools have been put together and the work has been done alone.
The change to a reminder of priority assignment is better, but on one occasion it was decided that things would be very small and that it would be long. prompt tells him to learn the division of labour, and he's working, and he thinks it's a small thing to call people?
The DeepSeek design divides answering and executing directly into different functions and does not place the division of labour in the decision-making of the model on the spot.
It is trying to calculate with more backstage calculations in exchange for shorter, more stable waiting and less human care.
For a person, it's worth it to see whether you save an occasional count or are interrupted all day.
Can you use it now? Try it, but don't open the blind box with the main computer.
The open warehouse, known as dsh-suberagent-prototype, currently operates around a GitHub code warehouse, with access to GitHub or flybooks. It is better suited to existing code projects, programmed Agent, and willing to handle human trials of deployment and permission, rather than a General Life Assistant ready for everyone.
To try, DeepSeek Harness needs to be configured, and the development environment required by the document. The prototype operated on the machine, and the shutdown could not count on it to continue drying, and the model had to be called and operated permanently at its own cost.
The official installation entrance was in line with its thinking, and the warehouse was handed over to Agent for programming, so AI could take you to install AI. You can send the following directly:
# Mission
Search for Deepseek-ai/dsh-suberagent-prototype in GitHub to read SETUP.md from the warehouse and take me to configure dsh-suberagent by document.
# Enforcement requirements
Pre-nuclear environment, collect the necessary information, give a complete plan and then install it after I have confirmed it.
As officially recommended, the active task begins with an observe and the code is consolidated by a person.
When you need an account login or a key, give me a local operation step. Do not let me send the key in the chat.
To determine which warehouse it works, which stand-alone account number, who is able to deliver, allow the list to be shorter and the code to be consolidated first.
Active missions are tacitly observed and not carried out directly.
It's gonna write down what it wants to take care of, and you're gonna see these judgments and decide whether to open it for automatic execution.
Allowing people on the list to clearly deliver is another path that can be implemented under observation.
This distinction is important, as the news is only that it is known and that there are requests for permission to start a new job or pre-written rules for the hit team. Otherwise you will complain, and it will immediately send its own life to save itself and turn it around.
permission also had to look at another layer, a prototype that had its system account number permission, which operated at the time of the execution of the order, and the task catalogue and documented packaging did not amount to safe isolation.
Official documents also recommend the use of a special system account number to test the environment in isolation, stand-alone accounts and unsensitized items, rather than handing over the documents on the main computer to the account number.
After configuration, the minimum attempt given by the document is to allow the account number in the list to be placed in the warehouse Issue @Agent, to send a small task with a clear range, to go to the panel to see if the task has occurred and to check for the changes it returns.
Let's not just look at how enthusiastic it is, change the request, see if it can catch it, meet the card points, see if it speaks. When it's not called, let's see if it calms down.
At the end.
Kim felt that it was not the addition of a Super in front of Agent that was of concern this time, but the attention of the people was finally taken as a cost to be carefully saved.
I used to cut back the window because it stopped. If it comes to you in the future, it'll have done what it can find out, only the one step that really needs your board, and this break is worth it.
So if you look at these tools again, you can ask more than how many things it can do at the same time. How many times do I have to come back after these things are handed over?
Of course, AI's busy. Don't be too busy until the end.
Thank you for reading my article.
If you think it's good, you can order it, read it, forward it.
If you want to get a delivery first, you can give me a sign. Thank you for reading my article.
Open-source knowledge base (real-time update exchange group):
my.feishu.cn/wiki/OhQ8wq…
Auto-fetched and translated by AI Atlas. All rights belong to the original author.