新智元报道
最近，DeepSeek又来炸场了。
昨天，DeepSeek公布了下一代Agent Harness的理念——Super-agent。
这篇长文指明：未来Agent的终极形态，将彻底打破「聊天窗口」的束缚，走向跨会话记忆、多任务并发的「真·Jarvis」时代。
文章一出，业界沸腾。
许多开发者直呼：「这就是我们一直在苦苦寻找的下一代AI应用形态！」
然而，当我们在为DeepSeek的顶层架构思维拍案叫绝时，居然发现了另一个事实——
早在今年4月，一个名为AFK AI（开源组织AFK-surf，产品名Comma）的硅谷华人团队，就已经在工程上完美实现了DeepSeek今天所描绘的几乎所有核心架构，并在时间上领先了全球几个月！
所以，当行业巨头还在发布「理念」时，这家神秘的团队是如何把「单 Session + 多 Worker」的超级架构变成现实的？
他们又将如何把超级智能带给每个人？
从此，再也没有多余的应用
很多人都发现，用AI的时候经常崩溃：比如聊着代码，忽然想让它回一封邮件。
为了防止上下文污染，你必须手动建个新对话；等回完邮件切回来，你又要重新梳理上下文。
据统计，人类的大脑最多只能同时跟进3到5个活跃的Session。传统的AI产品，其能力的上限不是模型的智商，而是人类的注意力。
在最新的文章中，DeepSeek一针见血地指出：现有的Agent设计无法支撑多任务和团队协作，必须打破Agent与单次手动会话的绑定！
系统应该是一个能够跨会话保持上下文、同时推进多项复杂任务且互不干扰的连续实体（Super-agent）。
令人震惊的是，这与AFK团队在今年4月13日内部定下的核心方向如出一辙。
AFK团队当时抛出了一个振聋发聩的理念：「一个团队，一个对话」（Single session + Multi-worker）。
他们认为，用户逐渐用IM（即时通讯）接入AI后，必然会变成主要使用「单Session」。不管是你、你的同事，还是在飞书、Slack、微信里的任何一句话，都应该汇聚到一个拥有长期记忆的「主脑」那里。
DeepSeek把这个痛点总结为理论，而AFK早已经在几个月前把代码提交到了GitHub，并做成了极其成熟的产品 Comma。
比如DeepSeek讲到的自动PR、自动观测跨端执行任务，AFK已经跑了大半年了。
如果你把DeepSeek的长文和AFK-surf的技术文档放在一起对比，你会产生一种「互相抄作业」的错觉。
但这其实是因为，相似的设计约束，必然导出相似的架构选择。 只有真正触碰到了当前Agent落地的天花板，才会不约而同地进化出这样的系统。
我们来看看他们令人惊叹的技术相似之处：
1. 双层架构：Router（主脑） + Worker（四肢）
为了解决「既要记住所有事，又要同时干很多事」的矛盾，两者都抛弃了「一个大模型干到底」的思路，采用了经典的双层架构。
DeepSeek设计了常驻的 Router 层和上下文隔离的 Worker 层。
其中，Router负责统揽全局、承接人类交互与分派任务，但不亲自干脏活累活；具体的任务发给独立的Worker，Worker在隔离的上下文中疯狂干活，干完再向Router汇报。
而AFK Comma 的解法完全一致！AFK将其定义为 Lead Agent (Router) + Subagents (Workers)。
用户或团队永远只跟那个长期在线的Router说话。Router就像一个人类主管，接到任务后，给对应的Worker开一个新Session。
Worker有自己的环境、提示词和工具，多个Worker并行运行互不干扰，繁杂的执行细节全留在Worker那边，只有结果才会汇总给Router。
Router有一个长期存在的主上下文Single session（单 session），代表「一个人」或「一个团队」，持续好几个月。
它记得之前发生过什么，也知道现在有哪些事在进行。Worker 是它派出去干活的手，而它自己就是那个一直在场、负责到底的「主人」。这就解决了「人和 AI 之间如何长期相处」。
这种架构带来了革命性的产品体验——
你不用再做「群主」去拉群，不用每次给AI设定人设。在Comma里，你只要说话就行了，剩下的路由和分发，Router会全权搞定。
2. 「灵魂出窍」：计算与执行环境的分离（BYOM）
普通的AI只能在云端的虚拟沙箱里执行任务，一旦涉及到公司内网代码、你本地的私人文件，AI就成了瞎子。
DeepSeek 提出，计算、状态与环境必须分离。
Agent的「脑」在云端，但「手」可以伸向用户的本地机器。通过在端侧部署一个不保存状态的轻量Daemon，云端Agent可以按需读取本地环境，这被称为BYOM（Bring Your Own Machine）。
而AFK Comma 的落地，再次领先一步——Comma已经把这个理念做成了强大的企业级功能！
通过在企业 IM 中 @Comma bot，Agent 可以请求员工终端执行任务。员工个人的 Personal agent 可以接管个人的电脑作为执行任务的节点。同时，企业中有人 @Comma bot 时，会触发 Agent 之间的调度与通信机制，继承人的权限与角色，在不牺牲信息安全的情况下，让 AI 替人协作并上班。
同时，这也意味着，Agent 组成的网络可以洞察企业细微的业务细节信息，并实现持续学习。
同时，系统内置了信息流控制的安全防护机制(ifc)，对敏感数据的访问与传输进行管控，保障信息安全。
这就相当于，DeepSeek刚刚论证了「脑手分离」的可行性，而AFK Comma已经把这套「云脑+本地手」的系统卖给企业用了！
3. 事件驱动与多端并发：真正的「全双工」
人类的工作不是「一问一答」的，而是异步的。
为此，DeepSeek 提出，引入事件驱动（Event-driven）和全双工交互（异步ReAct Loop）。模型在后台跑任务，前台依然可以秒回你的新消息。
而AFK Comma的客户端，直接打通了Telegram、Signal、微信、iMessage、飞书、Slack！
zanweiguo 直言：从此，啥App都不用开了。
它是一个真正的「工作站」。你在飞书里交代的事情，Router在后台派Worker去干；同事在Slack里追问进度，Router已经能同步回答。
甚至，从这些各大平台产生的Task，全都在Comma客户端统一管理；反过来，Comma也可以去第三方平台创建和管理Task。
这，就是真正的Jarvis。它无处不在，多端记忆绝对一致，不被任何一个App绑定，却能掌控所有的信息流。
以下，就是Comma的更多惊艳demo。
它可以帮你管理Slack。
它还自己修自己:
古风3D麻将，它也能轻松搞定。
甚至还能一起玩儿3D麻将，绝了！
它可以帮你盯机票。
路上随手拍一张iPhone DUO的广告牌，让Comma做出个复古滤镜app，一键梦回2012。
还能管理其他平台，比如Linear，可以帮你详细安排日程：
让它查局域网、看邮件、总结最近的新闻，通通搞定：
甚至可以让它修视频、拍视频。
当 DeepSeek 刚刚把 Super-agent 作为下一代方向抛出时，AFK AI 团队已经在工程实践的泥坑里摸爬滚打了大半年。
翻开 AFK AI 团队公开的工程记录和代码 Commit 历史，你会看到一条清晰的探索之路。
早在今年上半年，他们就已经把「单 Session + 多 Worker」架构推向了生产环境。
还在2025年9月时，硅谷出现了个人助理产品Poke，证明了「一个人，一条连续对话」的模式是成立的。但它仅仅局限于个人。
今年4月13日，团队内部拍板，敲定了「日常聊天的主 Agent + 干活的 Agent」的底层设计。
今年4月28日，Comma 的前身 Bridge 公开发布，博文明确写道：「Bridge is a lead agent... sub-agents are integrated automatically.」
今年5月20日，是历史性的一刻。他们在代码层面正式确立了「团队级单 Session」。每个团队拥有了一个稳定的 Router Session。
今年6月3日，固定的Bridge对话和 delegate_task 正式上线。用户只跟Router说话，Router全自动把任务派给Worker。
而现在，在今年10月底，DeepSeek 发布 Super-agent 架构长文，全行业开始瞩目这一方向。
复盘整个时间线，你会发现AFK AI团队对技术趋势的嗅觉极其敏锐。
注：插个有趣的八卦，AFK AI 最初的产品形态叫 cue，结果没想到Manus发布了类似名字的东西，因此，团队才改名叫 Comma。
而作为对比，Anthropic直到 今年 6 月 23 日，才在 Slack 推出类似理念的 Claude Tag。
在这个路线上，Comma 的公开发布比 Claude Tag 早了将近 2 个月，而底层架构的落地则早了整整 5 周！
比起DeepSeek 的理论公开，他们更是早了半年就完成了工程落地。
不仅如此，AFK-surf 坚持了开源、安全、支持自部署、支持BYOK。你可以自由地接入你自己的 Codex 或是 Claude，而 Comma Agent 能监督你的 Codex 和 Claude 干活。
这种开放性和架构的前瞻性，让 Comma 已经是下一代AI操作系统。
而且，DeepSeek在文章提到——「要突破限制，需要打破 Agent 与单次手动会话的绑定……并在长期运行中遇到异常时自动恢复。」
这些坑，Comma早就踩过，还给出硬核的解法，比如按任务隔离的 Worker，绝不污染主脑；系统级的信息流隔离；原子级的 Session 切换与 TLA+ 建模登顶等等。
这些，是真正把计算机系统的经典设计思想融入到了 Agent 架构中。
读懂了AFK Comma的底层逻辑，再结合DeepSeek的长文，我们其实已经看到了未来3到5年AI Agent的发展终局。
趋势一：单点Session的消亡，「一团队一主脑」成为标配
未来，你的公司、你的生活，都将只会拥有「一个」超级助理（Router）。所有成员、所有App（微信、Slack、飞书）的信息全盘汇入这个主脑。
趋势二：AI将从「被动响应」模型变为「主动推进」模型
基于双层架构和后台静默Worker的设计，未来的AI会持续观察公共群聊、GitHub、邮件，成为一个主动推进者。
趋势三：云端大脑 ，万物皆可Agent
由于环境和计算分离，未来的执行节点会下沉到每一个员工的电脑、每一台手机甚至每一台智能家居上。AI的脑子在云端，用「手」来操控本地资源。
昨天，DeepSeek用一篇长文告诉我们，对于未来的超级智能体，底层系统的革命才是真正的关键。
令人兴奋的是，在这个关键的转折点，这次在硅谷的华人团队跑在了最前面。
编辑：大卫 Aeneas
不喜欢
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。
New Ji-won.
Recently, DeepSeek has come to blow up again.
Yesterday, DeepSeek published the idea of the next generation of Agent Harness, Super-agent.
This long article points out that the ultimate form of the future of Agent will be to break the “chat window” completely and move on to the “Jarvis” era of multitasking memory and multitasking.
Article one, industry boils.
Many developers have been saying, "This is the next generation of AI applications we've been searching for!"
And yet, when we're making a case for DeepSeek's top structure, we find another fact.
As early as April this year, a Silicon Valley Chinese team called AFK-Surf (Open Source Organization) Coma has perfected almost all the core structures described today by DeepSeek and has been ahead of the world for months!
So, when the industry giant was still in publish, how did this mysterious team turn the "singion + more worker" superstructure into reality?
And how will they bring super intelligence to everyone?
Since then, there have been no more applications.
A lot of people find it very common to crash using AI: talking about code, for example, suddenly trying to get it back to an e-mail.
In order to prevent context contamination, you have to build a new conversation manually; when you get back to the mail, you have to recombine context.
According to statistics, the human brain can only follow three to five active sessions at the same time. The upper limit of the capacity of traditional AI products is not the IQ of the model, but the human attention.
In a recent post, DeepSeek pointed out with blood: the existing Agent design cannot support multitasking and teamwork, and the binding of Agent to a single manual conversation must be broken!
The system should be a continuous entity (Super-agent) capable of maintaining context over sessions while advancing complex tasks without interference.
It is alarming that this coincides with the core direction set by the AFK team internally on 13 April this year.
The AFK team then threw out a deafening idea: "One team, one dialogue."
They believe that the gradual use of IM (immediate communication) to access AI will inevitably become the main use of "single-session". Whether it's you, your colleagues, or any of the words in Flying Book, Slack, WeChat, should be brought together in a "master brain" with long-term memory.
DeepSeek summed up this pain point into theory, and AFK had submitted the code to GitHub a few months ago and made it an extremely mature product, Comma.
For example, when DeepSeek spoke of an automatic PR, an automatic observation cross mission, AFK has been running for over six months.
If you compare DeepSeek's long text with AFK-surf's technical files, you have the illusion of "blowing jobs."
But this is because similar design constraints necessarily lead to similar structural choices. It is only when there is a real touch on the ceiling of the current Agent that such a system will evolve in the same direction.
Let's look at their amazing technological similarities:
1. Double layer structure: Router + Worker
In order to resolve the contradiction between "recognizing everything and doing everything at the same time", both of them abandoned the idea of "a LLM to the end" and adopted the classic two-tiered structure.
DeepSeek designed the permanent Router layer and the context isolated Worker layer.
Among them, Router is responsible for managing the whole picture, carrying out human interactions and assignments, but not doing dirty work in person; specific tasks are given to independent Worker, who works crazyly in a segregated context and then reports back to Router.
The solution for AFK Comma is perfectly consistent! AFK defines it as the Head Age (Router)+ Subagents (Workers).
Users or teams always only talk to Router, who's been online for a long time. Router's like a human director, and when he gets the assignment, he's giving the corresponding Worker a new session.
Worker has its own environment, prompt and tools, multiple Worker parallel operations do not interfere with each other, and complex implementation details are left at Worker ' s side, and only the results will be aggregated to Router.
Router has a long-standing master contextSingle session (single session), representing " one person " or " a team ", which lasts for months.
It remembers what happened before and knows what is going on. Worker is the man he sent to work, and he's the one who's always there and is responsible for the whole thing. This solves the problem of "how people and AI live together for a long time."
This structure brings a revolutionary product experience -
You don't have to be the master of the pack, you don't have to set it up for AI. In Coma, you just talk, the rest of the route and distribution, and Router will take care of it.
2. "Spirit of the soul": the separation of calculations from the enforcement environment (BYOM)
Normal AI can only operate in a cloudy virtual sandbox, and once it involves the company's intranet code, your local personal files, it becomes blind.
DeepSeek argues that computation, state and environment must be separated.
Agent's "brain" is in the clouds, but the "hand" can reach the user's local machine. By deploying a light, unsaved-state Damon on the end, the cloud-end Agent can read the local environment on demand, known as BYOM (Bring Your Own Machine).
And the landing of AFK Comma is another step ahead -- Comma has made it a powerful enterprise-level function!
Through @Comma bot, Agent can request an employee terminal to perform the task. Personal agent can take over the personal computer as the node of the mission. At the same time, when someone in the enterprise @Comma bot triggers a movement and communication mechanism between Agent, the successor permission and his character, who, without sacrificing the security of the information, allows AI to collaborate and work.
At the same time, this means that Agent ' s network can provide insight into the fine details of business and achieve continuous learning.
At the same time, a security protection mechanism (ifc) for flow control has been put in place within the system to control access to and transmission of sensitive data and to secure information.
This means that DeepSeek has just demonstrated the feasibility of "separation of the brain" and AFK Comma has sold the system to the company!
3. Event-driven and multi-pronged: a true "full-time job"
The work of human beings is not "a question-and-answer"; it's a walk.
To this end, DeepSeek proposes the introduction of Event-driving and full-work interaction (otherwise rect Loop). The model runs in the backstage, and the front desk can still return your news.
And AFK Comma's client, who has direct access to Telegram, Signal, WeChat, iMessage, Flying Book, Slack!
zanweiguo: From now on, there's no need to turn on anything.
It's a real "workstation". What you said in the flying book, Router sent Walker backstage to do it; colleagues asked for progress in the Slack, and Router was able to answer simultaneously.
Even Task, which emerges from these major platforms, is managed centrally by Comma clients; in turn, Comma can go to a third-party platform to create and manage Task.
This, this is the real Jarvis. It is ubiquitous, multi-end memories are absolutely consistent, not bound by any App, but capable of controlling all information flows.
Here's Comma's more amazing demo.
It can help you manage Slack.
It also fixes itself:
Old 3D mahjong, it can be easily handled.
We can even play mahjong 3D together.
It'll help you with the ticket.
On the way, an iPhone DUO billboard was taken, allowing Comma to make a retrospect and dream back to 2012.
Other platforms, such as Linear, can help you to schedule things:
Let it check the local area network, read the mail, summarize the latest news, and get it done:
It can even fix the video, make video.
When DeepSeek just threw Super-agent in the next generation, the AFK AI team had been crawling in the mudhole of engineering practice for six months.
Turning the AFK AI team's public project records and code Committee history, you'll see a clear path to exploration.
As early as the first half of this year, they have moved the "Solution + More Worker" structure to a production environment.
Also in September 2025, Poke, a personal assistant product, appeared in Silicon Valley, proving that the “one person, one continuous dialogue” model had been established. But it's just personal.
On April 13, this year, the team board finalized the bottom-up design for the "Agent + Worker of Daily Chat".
On April 28th of this year, Bridge, the predecessor of Comma, published publish, wrote clearly in Bo: "Bridge is a lead agent... sub-agets are integrated obviously."
This 20 May is a historic moment. They officially created the "team-level session" at the code level. Each team has a stable Router Session.
On 3 June this year, the regular Bridge Dialogue and delegate task officially went online. Users only talk to Router, who automatically assigns tasks to Walker.
And now, at the end of October of this year, DeepSeek publish Super-Agent, a long-form architecture, is beginning to gain industry attention in this direction.
Rewind the entire time line and you'll find that the AFK AI team is extremely sensitive to technological trends.
Note: In an interesting gossip, AFK AI's original product form was Cue, but it didn't come to light that Manus publish had something like that, so the team changed to Comma.
And by contrast, Anthropic did not introduce a similar idea of Claude Tag in Slack until June 23rd this year.
On this route, Comma's public publish is nearly two months ahead of Claude Tag, and the bottom structure is five weeks ahead of ground!
They landed six months earlier than DeepSeek's theory.
In addition, AFK-surf insisted on open sources, security, support for self-deployment, and support for BYOK. You have free access to your own Cordex or Claude, and Comma Agent can oversee your Cordex and Claude working.
This openness and the forward-looking architecture make Comma a next-generation AI operating system.
Moreover, DeepSeek mentioned in his article that "to break the limit, it is necessary to break Agent's binding on a single manual conversation... and to recover automatically in the event of an anomaly in a long run."
These pits, which Comma stepped on long ago, also gave hard core solutions, such as task-segregated Walker, which does not pollute the main brain; system-level information flow isolation; atomic-grade Ssession toggle to the top of the TLA+ model.
And this is really the idea of incorporating the classic design of computer systems into the Agent architecture.
We've read the bottom logic of AFK Comma, combined with DeepSeek, and we've actually seen the final development of AI Agent over the next three to five years.
Trends one: the demise of a single-point session, the "one team, one brain" becomes the standard
In the future, your company and your life will only have a "one" super assistant. All members, all Apps (WeChat, Slack, Flying Books) reciprocated into the brain.
Trends II: AI will move from a "passive response" model to a "active propulsion" model
Based on the design of a two-tiered structure and a quiet Worker in the back, the future AI will continue to observe public conversations, GitHub and mail as a proactive pusher.
Trends three: cloud brain, everything is Agent.
Because of the isolation of the environment and calculations, the future executive nodes will sink onto every employee ' s computer, every cell phone and even every smart house. AI's brain is cloudy, using "hands" to manipulate local resources.
Yesterday, DeepSeek told us in a long article that the revolution of the bottom system is the real key to the future super-AI agent.
It is exciting to note that, at this critical turning point, the Chinese team in Silicon Valley was at the forefront this time.
Editor: David Aeneas
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.