文 | 字母AI
DeepSeek Harness团队在10月31日晚上发布了一篇文章，标题为《Super-agent设计探索：迈向更自主的 Agent 系统》，文章里提出了这样一个概念：Super Agent。
同时他们也在GitHub上开源了Super Agent的原型dsh-superagent-prototype，以供开发者参考。
所谓Super Agent，它就相当于一个记忆力超群、24小时在线、永远不摸鱼偷懒的同事。
你之前交代过什么，即使到下一个任务里，它依然能够记得。还能在各种聊天软件、邮件和GitHub等不同的地方，执行同一项工作。
“主动”是Super Agent的核心关键词，有些事情不一定非得等到你开口命令它，它才会去干活。
比如看到群里有人报Bug，哪怕没人艾特Super Agent，它也能先去看看问题出在哪，能修就直接动手。
Anthropic此前推出的Claude Duet，虽然还处于预览版，但是它也会主动看邮件和日历，早上给你一份简报，平时发现重要事情再来找你。
Agent不该一直等你开口，Super Agent和Duet的共同点都是如此。
但“主动”并不意味着刷存在感。
一个真正好用的Super Agent，不仅要知道什么时候该主动找你，更要知道哪些事可以自己处理，什么情况才允许打扰你，什么时候应该保持安静。
它知道该干什么，也知道什么时候别烦你。
Super Agent和普通Agent有啥区别？
Super Agent和普通Agent的区别就在于，前者像是正式员工，后者更像是一个临时工。
你交代给普通Agent一件事，它会开始查资料、写代码、给结果。
当你下次有别的事，即使你叫来的还是它，它又要从查资料开始干起。就算普通Agent把所有事情都经历一遍，可是到了最后，还是只有你自己知道所有任务的全貌。
因为每个任务都是一个独立的会话（Session），Session之间互不相通。普通Agent在A Session里得出的结论，B Session里的他根本不知道。只有你自己才知道整个任务的全貌。
可DSH团队观察到，一个人的精力有限，通常只能同时跟进3到5个这样的活跃Session。所以问题不只是Agent够不够聪明，你能同时盯住几个Agent也是提升模型工作效率的重点。
那如果让一个员工同时做所有事呢？
DSH认为这样也不现实。普通Agent一边修Bug，一边写报告，还要不断回消息，输出速度必然会变慢，更可怕的是几件任务的细节也容易混在一起，导致最后的结果驴唇不对马嘴。
Super Agent要解决的第一件事，就是分配两种角色。
一种叫Router，可以理解为项目主管。他负责听需求、记背景、看全局，再决定任务交给谁，但不亲自去做具体执行。
另一种叫Worker，就是真正干活的员工。每个Worker只领一项任务，在自己的独立空间里完成，修Bug的专心修Bug，写报告的专心写报告，几个人可以同时开工。
有句话听起来挺扯的，可事实就是这些分工不同的Worker，它们的“人格”也是不同的。有些好沟通，有些极难沟通。
所以在Super Agent这里，你只需要跟最好沟通的项目主管打交道，不必挨个跟每个Worker沟通。
看到这里，你可能会问：这不就是主Agent带着几个子agent吗？前几天Anthropic刚刚发文章，标题叫《构建Agent》，核心就是说弄一堆子Agent不如单个Agent效率高。
事实上两者截然不同。
负责统领子Agent的主Agent，它的任务非常繁重，包括深入理解任务细节、规划子任务、高频协调，所以它的推理负担很重。同时子Agent每一步都需要耦合，所以就需要主Agent频繁协调。
Router根本不考虑那些，它只管“这活我应该推给谁”，Worker之间上下文隔离、彼此独立，靠订阅同一个Thread感知冲突。
所以从产品逻辑上来说，Super Agent跟子Agent完全是两码事。
Super Agent的“主动”从何而来？
Super Agent干的第二件事是，没人专门下命令时，它怎么主动去干活？
DeepSeek的做法让它留意工作环境里的动静。
任何动静都可能成为一个“触发器”（Trigger），Super Agent一旦察觉有动静了，就会立刻进入工作状态。
任务派出去之后，后续沟通也不必一直经过主管。比如Worker已经开始修某个Bug，那么这个话题下面的新评论，就会直达Worker。
就像维修师傅接单以后，你会直接把问题报给维修师傅，不用再让客服转述一遍。
DSH的文章中有个案例，开头也提到了，群里有人说：“这个项目有BUG。”没人艾特Agent，但这句话进入系统后，先经过规则过滤，再由模型判断它符合团队提前写好的“静默复现并修复”策略。
项目主管把任务交给一个Worker，Worker在自己的独立空间里复现问题、查找原因、尝试修复。没复现出来，或者发现修复很复杂，它就保持安静。如果确认问题并找到简单方案，就提交一份PR，最后交给人验收结果。
后续有人在GitHub评论里要求换个修法，Worker也能从原来在聊天软件的场景里换到GitHub上面接着做。
DeepSeek文章里展示的内部实例，就是这样在群聊和GitHub之间连续推进任务，甚至在无人艾特的情况下主动提交修复代码的申请。
除了看聊天软件里的消息以外，Super Agent还会观察所有能看到的信息，比如收到一封新邮件、日程有变化、代码自动检查没通过。
但是群聊里大部分消息都是闲聊，绝大多数跟Super Agent没关系。如果每条都让模型读一遍，又慢又贵。
所以DeepSeek给Super Agent加了两道筛网。
第一道是提前定好的规则：哪些群能看、哪些人或机器人的消息算数、哪些来源直接忽略。
第二道交给模型判断：剩下的候选消息，这句话是随口吐槽，还是真有一个需要处理的问题？
这类需要理解语义的判断，规则没办法写出来。好在的是，第二道筛网是用自然语言写的，想改策略，改一句提示词就行。
只有两道都通过了，项目主管才会派任务。
Super Agent另一大特点是把工作放到后台，先告诉你“已经开始做了”，做完再把结果送回来。这期间，你仍然可以补充要求、叫停，或者安排下一件事。
DeepSeek称，这一异步任务机制已经在DSH中实现，但尚未默认开启。
你去餐厅点菜，点完之后服务员也不可能站在灶台前等菜炒完，他会回来继续招呼其他客人，等菜做好了他才去端上来给你。
还有一个事，DeepSeek认为Agent是要持续干活的，所以Super Agent更是要一直在线、持续工作。可是干活久了，上下文就会太长，Agent很容易出错。
因此DeepSeek把Agent的“大脑”、任务记录和实际操作的电脑分开。
大脑可以在云端工作，任务记录单独保存，代码和文件仍留在原来的电脑上。这样，某台云端机器重启了，记录还在。需要操作你的电脑时，再远程连接过去。
同时，DeepSeek还给Super Agent安排了一个Watchdog（看门狗）。
普通的监控只能看Agent有没有宕机，Watchdog还要看任务有没有进展、方向有没有跑偏。
就像店长巡店，不只是看灶台的灯亮没亮，还要看厨师放了几克盐、煮了多长时间。
不过，GitHub上的dsh-superagent-prototype还不能被当成Super Agent的“完全体”。DeepSeek表示，它只是研究原型，不是正式产品。目前主要面向一个GitHub仓库，也可以选择接入飞书，文章里的一些设计还没有实现。
Super Agent最强的能力是“闭嘴”
GitHub原型的说明文档里，有这么一句话：尽可能少占用人的注意力，尽可能多地提供价值。
DeepSeek把每一条消息、每一次通知、每一个问题、每一个设置项，都看成一种注意力成本。不能提供足够价值，就应该删掉。
原型还明确规定，Agent工作时保持安静，只有结果可用、需要人做决定，或者任务无法完成时，才再次开口。
这跟过去互联网产品的逻辑几乎是反着来的。
以前做产品，往往希望你多打开。Super Agent要做的是帮你把事情处理掉，让你少打开。
一个真正好用的Agent，不应该把20条群消息总结成另外20条通知，再让你逐个确认。它应该先把大部分事情做完，你只需要做那个非你不可的决定。
所以，Super Agent主动性的核心不是“多说话”，是它知道什么时候不该打扰你。
DeepSeek文章里有个“静默复现并修复”的例子。
Agent可以先在后台复现Bug、寻找修复方案，但只有确认问题并找到简单方案后才交付结果，否则全程保持安静。
换成一个普通Agent，可能刚看到消息就先回一句“收到，我正在分析”，十分钟后发来一段过程，半小时后再告诉你没复现。
是，这个普通Agent很忙，我看着都累，可你也确实被它折腾了半天。
可Agent产生的信息，其实不等于Agent提供的价值。Super Agent就是在重新定义人跟AI之间的关系，人不应该给AI当秘书。
DeepSeek反复强调一个概念，Session该由系统在背后管理，用户不应该为它操心。
文章提到“从这个角度看，让用户手动管理Session，在很大程度上是从底层LLM API延续下来的设计惯性”。这句话的意思是，会话并不是正确的产品形态，它只是早期大模型API留下的习惯。
既然用户真正想管理的是任务，那就应该让系统自己去管会话，人只需要看事情有没有办完。
有了Super Agent，你就不用每天特意打开某个AI应用，它只要在群里、代码仓库和邮件里，就能自己工作。
当然，Agent越主动，它的边界感也就越重要。
DeepSeek的Super Agent有这样一条规则：Super Agent只在被请求时，或者书面规则允许时行动。没有人直接要求的任务，需要团队预先写下规则。
这就像你新招了一个员工，刚开始，你让他先写方案，你得先过一遍。只有确认他理解任务以后，才慢慢放权给他。哪些结果要审批、哪些操作必须停下来问人，都得提前讲清楚。
Anthropic的Duet也是同样的思路。它可以自己读信息、起草回复、准备工作，但发送消息、付款、签署文件、删除内容之前，必须等人批准。
它每天早上主动提供简报，每隔几小时检查一次，避免后台任务把额度快速烧完。
“闭嘴”不是能力不够，恰恰是一种更高级的产品能力。它是一种克制，没有结果时不播报过程，没有权限时不擅自行动，没有必要时不制造新消息。
而克制，也是梁文锋的关键词。在此前的投资者交流会上，梁文锋说：“你越克制，越有可能做成AGI。”
不过有一点必须要提醒。Super Agent是在用更多的token，换人的时间与注意力。
DeepSeek表示，多Session设计会带来更多token消耗。Duet的晨报和后台检查，同样会占用用户的套餐额度。
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。
Letter AI
The DeepSeek Harness team presented an article entitled " Super-agent Design Exploration: Towards a More Autonomous Agent System ", entitled publish, on the evening of 31 October, which introduced the concept: SuperAgent.
At the same time, they have also launched the prototype of SuperAgent, dsh-suberagent-prototype, on GitHub, for the information of developers.
So-called SuperAgent, it is equivalent to a 24-hour online colleague who has memory superstitions and never touches a lazy fish.
What have you said before, even in the next mission, it will remember. The same work can also be carried out in different places, such as chat software, mail and GitHub.
“Initiative” is the core keyword of SuperAgent, and some things do not necessarily have to wait until you ask for it.
See, for example, there are people in the group reporting Bug, and even if there's no one named Et Super Agen, he can go and see where the problem is, and fix it.
Claude Duet, previously launched by Anthropic, is still in the preview, but it also takes the initiative to look at mail and calendars, gives you a briefing in the morning and normally finds important things to come back to you.
Agent shouldn't have waited for you to say it. It's what Super Agent and Duet have in common.
But “active” does not mean a sense of presence.
A really good SuperAgent, not only knows when to come to you, but also what to deal with, what to disturb you and when to keep quiet.
It knows what to do and when to leave you alone.
What's the difference between Super Agent and normal Agent?
The difference between SuperAgent and normally Agent is that the former looks like a regular employee and the latter more like a temporary worker.
You tell normal Agent one thing, it starts searching for information, writing codes, giving results.
And when you do something else, even if you call it, it starts with the search. Even though ordinary Agent went through everything, at the end of the day, you were the only one who knew the full picture of everything.
Because each mission is an independent session, there is no connection between the two. Normal Agent came to the conclusion in A Session that he didn't know anything about B Session. Only you know the full picture of the mission.
The DSH team observed, however, that one person has limited energy and can usually follow only three to five such active sessions simultaneously. So the question is not just whether Agent is smart enough, but you can keep an eye on a couple of Agents at the same time.
What if an employee does everything at the same time?
DSH finds this unrealistic. The normal Agent, while repairing Bug while writing his report, keeps getting back to the news, and the speed of the output is bound to slow down, and even worse, the details of several missions are easily mixed, resulting in the end result being the wrong mouth.
The first thing Super Agent has to do is assign two roles.
One called Router, understood as the project manager. He is responsible for listening to needs, keeping a background, looking at the whole picture before deciding to whom to entrust the task, but not for carrying it out in person.
The other is Worker, the real worker. Each Worker is assigned only one mission, performed in his or her own independent space, to fix Bug's focus, to write his or her report, and several people can start work at the same time.
There is a saying that sounds crazy, but the fact is that these different divisions of work are different. Some are good and some are extremely difficult to communicate.
So here at Super Age, you're just gonna have to deal with the best project manager, and you're not gonna have to talk to each and every one of them.
Seeing here, you might ask, "Isn't this Lord Agent with a few children?" Anthropic just published a few days ago under the title "Building Agent ", the core of which is to make a pile of Agent less efficient than a single Agent.
In fact, the two are very different.
Agent, the head of the command of the co-lead, has a heavy task, including in-depth understanding of mission details, planning of subtask, and HF coordination, which places a heavy burden on its reasoning. At the same time, every step of Agent needs to be combined, so the main Agent is required to coordinate frequently.
Router doesn't give a shit about that, he just says, "Who am I supposed to give this job to?" context separates and separates from each other between Works by subscription to the same Thread perception conflict.
So in product logic, Super Agent and Child Agent are completely different things.
Where does Super Agent's initiative come from?
The second thing Super Agent does is how does it get to work when no one orders it?
DeepSeek's approach allows him to keep an eye on the movement in the work environment.
Any movement could become a Trigger, and SuperAgent would be operational as soon as it was detected.
Follow-up communication does not have to go through the supervisor after the mission. For example, Worker has begun to fix some Bug, so the new comments under this topic will go straight to Worker.
As soon as the maintenance master gets the order, you'll report the problem directly to the maintenance master without having to repeat it.
There is a case in the DSH article, which is also mentioned at the beginning, where some people say, “The project has a BUG.” There's no one called Agent, but after entering the system, the phrase is filtered through the rules, and then the model judges it to be in line with the “quiet and repair” strategy written earlier by the team.
The project manager assigned the task to a Worker, who reproduced the problem in his own independent space, sought out the reasons and tried to repair it. It's quiet when it doesn't come back, or when it's complicated. If the problem is identified and a simple solution is found, a PR is submitted and the result is finally presented to the person acceptance check.
Following a request in GitHub’s review, Worker could move from the original chat software scene to GitHub.
The internal example shown in DeepSeek's article is the way in which successive missions are carried out between the group conversations and GitHub, even with no one, Eid, who voluntarily submitted an application for the repair code.
In addition to reading the news from the chat software, SuperAgent will observe all the information that can be seen, such as a new e-mail, a change in schedule, and an automatic code check.
But most of the information in the group is chatty, and most of it has nothing to do with Super Agent. If you read every one of them, it's slow and expensive.
So DeepSeek added two screens to Super Agent.
The first is a pre-established rule: which groups can see, which people or robots can count, which sources can ignore directly.
The second is for the model to judge: the rest of the candidate's information, is the sentence free-mouth, or is there really a problem that needs to be addressed?
This kind of judgement requires understanding of semantics, and rules cannot be written. Fortunately, the second sift was written in natural language, so we could change our strategy and change prompt.
Only when both tracks are adopted will the project manager be assigned.
Another important feature of Super Agent is to put the job backstage and tell you “do it already”, and then bring it back. In the meantime, you can still ask for additional, stop or arrange the next thing.
DeepSeek states that this different mission mechanism has been achieved in DSH but has not yet been implicitly activated.
You can order at the restaurant, and the waiter can't wait until the cooker's ready. He'll come back and take care of the other guests. He'll bring them to you when they're ready.
Another thing is that DeepSeek believes that Agent has to work on a continuous basis, so SuperAgent has to work on a continuous basis. But after working long hours, context will be too long, and Agent is prone to mistakes.
So DeepSeek separates Agent's “brain”, mission records and operational computers.
The brain can work in clouds, mission records are kept separately, and codes and files remain on the original computer. So, a cloud machine is restarted and the records are still there. When you need to operate your computer, connect it remotely.
At the same time, DeepSeek also arranged a Watchdog for Super Agent.
The usual surveillance is to see if Agent is dead, and Watchdog is to see if the mission is moving or running.
Like the shopkeeper, not only did the stove lights turn on, but the cook put a few grams of salt and cooked it for how long.
However, dsh-suberagent-prototype on GitHub cannot yet be considered as the “full body” of SuperAgent. DeepSeek states that it is a research prototype and not a formal product. Currently, the main target is a GitHub warehouse, with the option of access to flying books, and some of the designs in the article have not yet been realized.
Super Agent's greatest ability is to shut up.
In the GitHub prototype description document, there is a sentence that minimizes the attention of the occupants and provides as much value as possible.
DeepSeek sees every message, every notice, every question, every setup as a cost of attention. Failure to provide sufficient value should be eliminated.
The prototype also specifies that Agent works quietly and speaks again only when the results are available, decisions are required or tasks cannot be accomplished.
This is almost the opposite of the logic of past Internet products.
Used to be a product, always wanted you to open more. Super Agent is gonna help you get rid of things and keep them open.
A really good Agent should not summarize 20 sets of messages into 20 other notifications before you confirm them one by one. It should finish most things first. You just have to make that decision.
So the core of SuperAgent's initiative is not "do more," it knows when to disturb you.
In DeepSeek's article, there is an example of a “silent return and repair”.
Agent could recreate Bug in the backstage and look for a rehabilitation programme, but only if he identified the problem and found a simple solution would the results be delivered, otherwise the whole process would remain quiet.
In the case of an ordinary Agent, the word “received, I'm analysing” may be answered immediately after the news, and a process will be sent in 10 minutes and you will be told in half an hour that it has not been repeated.
Yeah, this regular Agent is busy, and I look tired, but you've been in it for half a day.
But the information that Agent produces does not actually equal the value that Agent provides. SuperAgent was redefining the relationship between people and AI, and people should not have been secretary to AI.
DeepSeek has repeatedly emphasized the concept that Session should be managed by the system behind its back and that users should not worry about it.
The article mentions that “from this perspective, allowing users to manage Session manually is largely a continuation of design inertia from the bottom LLM API”. This sentence means that the session is not the right product form, but is a habit left by the early LLMAPI.
Since it is the task that the users really want to manage, the system should be left to manage its own sessions, and one needs only to see if it is done.
With Super Agent, you don't have to open an AI application every day. It works in a group, in a code warehouse and in mail.
Of course, the more Agent takes the initiative, the more important its sense of boundary.
DeepSeek ' s SuperAgent has a rule that SuperAgent acts only when requested or when written rules permit. The task for which no direct request has been made requires advance writing by the team.
It's like you've got a new employee, and you've got him to start with. You've got to go through it. Only after confirmation of his understanding of the mission would he be slowly decentralized. It must be clear in advance what results are to be approved and what operations must stop and ask.
The same is true of the Duet of Anthropic. It can read information itself, draft responses, prepare work, but before sending it, paying for it, signing documents and deleting content, approval must be sought.
It offers briefings every morning and checks every few hours to avoid the quick burnout of quota by the backstage mission.
“Shut up” is not a lack of capacity; it is a higher product capacity. It is a form of restraint, failure to broadcast the process when there is no result, lack of authority to act without permission and without creating new information when necessary.
And restraint is the key word for Liang Wenbei. At an earlier investor exchange, Liang Wenfeng said: “The more you restrain, the more likely you are to become AGI.”
But one must be reminded. Super Agent is using more token for time and attention.
DeepSeek stated that multi-session design would result in more token consumption. Duet ' s morning newspapers and back-office checks also occupy the user ' s quota.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.