DeepSeek刚发的下一代Agent，竟被这家硅谷团队提前半年实现了?
The next-generation Agent just released by DeepSeek was actually implemented six months early by this Silicon Valley team?
Content Summary
Xinzhiyuan reports Recently, DeepSeek caused a stir again. Yesterday, DeepSeek announced the concept of the next-generation Agent Harness — Super-agent. This long article points out:
Synced Report
Synced Report
Recently, DeepSeek caused a stir again.
Yesterday, DeepSeek announced the concept of the next‑generation Agent Harness—Super‑agent.
This long article points out: The ultimate form of future agents will completely break the constraints of the “chat window” and usher in a “true Jarvis” era of cross‑session memory and multi‑task concurrency.
When the article was released, the industry boiled over.
Many developers exclaimed, “This is exactly the next‑generation AI application form we’ve been tirelessly searching for!”
However, while we were applauding DeepSeek’s top‑level architectural thinking, we actually discovered another fact—
As early as April this year, a Silicon Valley Chinese team called AFK AI (open‑source organization AFK‑surf, product name Comma) had already flawlessly implemented almost all of the core architecture that DeepSeek describes today, and was months ahead of the rest of the world!
So, while industry giants are still publishing “concepts,” how does this mysterious team turn the “single session + multiple workers” super architecture into reality?
How will they bring super intelligence to everyone?
From now on, there will be no redundant apps
Many people have noticed that using AI often crashes: for example, you’re discussing code and suddenly want it to draft an email.
To prevent context contamination, you have to manually start a new conversation; after finishing the email and switching back, you need to re‑establish the context.
According to statistics, the human brain can track at most 3 to 5 active sessions simultaneously.For traditional AI products, the ceiling of capability is not the model’s IQ but human attention.
In the latest article, DeepSeek incisively points out:The current Agent designs cannot support multitasking and team collaboration; the binding between an Agent and a single manual session must be broken!
The system should be a continuous entity (Super-agent) that can maintain context across sessions, advance multiple complex tasks simultaneously, and keep them independent of each other.
Shockingly, this aligns perfectly with the core direction the AFK team set internally on April 13 this year.
The AFK team then put forward a striking concept: “One team, one conversation” (Single session + Multi‑worker).
They believe that as users increasingly access AI via IM (instant messaging), they will inevitably primarily use a “single session”. Whether it’s you, your colleagues, or any message in Feishu, Slack, or WeChat, it should converge to a “master brain” with long‑term memory.
DeepSeek frames this pain point as a theory, while AFK submitted the code to GitHub months ago and turned it into the highly mature product Comma.
For example, the automated PR and cross‑platform task observation that DeepSeek mentions have been running in AFK for over half a year.
If you compare DeepSeek’s long-form content with AFK-surf’s technical documentation side by side, you get the illusion of “copying each other’s homework.”
But this is actually because similar design constraints inevitably lead to similar architectural choices. Only when you truly hit the ceiling of current Agent deployment do you independently evolve such a system.
Let’s look at their impressive technical similarities:
1. Dual-layer architecture: Router (brain) + Worker (limbs)
To resolve the contradiction of “having to remember everything while doing many things at once,” both abandon the “single large model does everything” approach and adopt a classic two‑layer architecture.
DeepSeek designs a persistent Router layer and a context‑isolated Worker layer.
The Router is responsible for overseeing the whole system, handling human interaction and dispatching tasks, but it does not do the dirty or heavy work itself; specific tasks are sent to independent Workers, which work intensively in isolated contexts and then report back to the Router.
The solution of AFK Comma is exactly the same! AFK defines it as Lead Agent (Router) + Subagents (Workers).
Users or teams always communicate only with the long‑online Router. The Router acts like a human supervisor: when it receives a task, it opens a new session for the appropriate Worker.
Each Worker has its own environment, prompts, and tools. Multiple Workers run in parallel without interfering with each other; all complex execution details stay with the Workers, and only the results are aggregated back to the Router.
The Router has a long‑standing primary context called Single session, representing “one person” or “one team,” lasting for several months.
It remembers what happened before and knows what is currently going on. The Worker is the hand it sends out to do the work, while it itself is the ever‑present “owner” who sees things through. This solves the problem of “how people and AI can coexist over the long term.”
This architecture delivers a revolutionary product experience—
You no longer need to act as a “group admin” to gather a group, nor set a persona for the AI each time. In Comma, you just speak, and the routing and distribution are handled entirely by the Router.
2. “Out‑of‑Body”: Separating Compute and Execution Environments (BYOM)
Typical AIs can only run tasks inside a cloud‑based virtual sandbox; as soon as they need to access internal corporate code or your private local files, the AI becomes blind.
DeepSeek argues that computation, state, and environment must be separated.
The agent’s “brain” resides in the cloud, but its “hands” can reach the user’s local machine. By deploying a stateless lightweight daemon on the client side, the cloud agent can read the local environment on demand—this is called BYOM (Bring Your Own Machine).
With the implementation of AFK Comma, we are once again a step ahead—Comma has turned this concept into a powerful enterprise‑grade feature!
By @Comma bot in the enterprise IM, an Agent can request an employee’s terminal to execute tasks. An employee’s personal Personal agent can take over their computer as a task execution node.At the same time, when someone in the enterprise @Comma bot, it triggers the scheduling and communication mechanisms among Agents, inheriting the user’s permissions and roles, allowing AI to collaborate and work on behalf of people without compromising information security.
At the same time, this also means that a network composed of agents can gain insight into the subtle business details of an enterprise and achieve continuous learning.
At the same time, the system includes a built‑in information flow control security mechanism (IFC) that governs access to and transmission of sensitive data, ensuring information security.
This is akin to DeepSeek just demonstrating the feasibility of “brain‑hand separation,” while AFK Comma has already sold this “cloud brain + local hand” system to enterprises!
3. Event‑driven and multi‑endpoint concurrency: true “full‑duplex”
Human work isn’t a simple “question‑and‑answer” exchange; it’s asynchronous.
To address this, DeepSeek proposes introducing event‑driven architecture and full‑duplex interaction (asynchronous ReAct Loop). The model runs tasks in the background while the front end can still reply to your new messages instantly.
And the AFK Comma client directly connects Telegram, Signal, WeChat, iMessage, Feishu, Slack!
zanweiguo bluntly says: From now on, you don’t need to open any app.
It’s a true “workstation.” Tasks you assign in Feishu are dispatched by the Router to Workers in the background; when a colleague asks for progress in Slack, the Router can already sync a response.
Even the tasks generated by these major platforms are all managed uniformly in the Comma client; conversely, Comma can also create and manage tasks on third‑party platforms.
This is the true Jarvis. It is everywhere, with perfectly consistent memory across devices, not bound to any single app, yet capable of controlling all information flows.
Below are more stunning demos of Comma.
It can help you manage Slack.
It also repairs itself:
Ancient‑style 3D mahjong, it can handle it with ease.
You can even play 3D mahjong together—awesome!
It can help you keep an eye on flight tickets.
Snap a photo of an iPhone DUO billboard on the road, let Comma create a retro filter app, and instantly travel back to 2012.
It can also manage other platforms, such as Linear,and can help you schedule in detail:
Let it scan the LAN, check emails, summarize recent news—everything handled:
It can even edit videos and shoot videos.
When DeepSeek first touted Super-agent as the next-generation direction, the AFK AI team had already been slogging through the engineering trenches for over half a year.
If you open the AFK AI team’s public engineering logs and code commit history, you’ll see a clear path of exploration.
As early as the first half of this year, they had already pushed the “single Session + multiple Worker” architecture into production.
In September 2025, a personal assistant product called Poke appeared in Silicon Valley, proving that the “one person, one continuous conversation” model works. However, it was limited to individuals.
On April 13 this year, the team internally approved and finalized the underlying design of “the primary Agent for everyday chat + the Agent for getting work done.”
On April 28 this year, Bridge, the predecessor of Comma, was publicly released, and the blog post clearly stated: “Bridge is a lead agent… sub-agents are integrated automatically.”
May 20 this year marked a historic moment. They formally established a “team‑level single Session” at the code level.Each team now has a stable Router Session.
This June 3rd, the fixed Bridge dialogue and delegate_task officially launched. Users only talk to the Router, and the Router automatically assigns tasks to Workers.
Now, at the end of October this year, DeepSeek released a comprehensive article on the Super-agent architecture, drawing industry-wide attention to this direction.
Looking back at the timeline, you’ll see that the AFK AI team has an extremely keen sense of technological trends.
Note: A fun tidbit—AFK AI’s original product was called cue, but when Manus released something with a similar name, the team renamed it to Comma.
By comparison, Anthropic only launched the similarly‑concepted Claude Tag on Slack on June 23 this year.
On this timeline, Comma’s public release came almost two months before Claude Tag, and its underlying architecture was deployed a full five weeks earlier!
Compared to DeepSeek’s theoretical release, they completed the engineering rollout half a year earlier.
Moreover, AFK-surf adheres to open source, security, self‑deployment support, and BYOK. You can freely integrate your own Codex or Claude, and the Comma Agent can supervise your Codex and Claude at work.
This openness and forward‑looking architecture already make Comma a next‑generation AI operating system.
Furthermore, DeepSeek mentions in the article—“To break limits, we need to decouple agents from single manual sessions… and automatically recover when anomalies occur during long‑running operations.”
These pitfalls have long been tackled by Comma, which offers robust solutions such as task‑isolated Workers that never pollute the main brain; system‑level information flow isolation; atomic Session switching and TLA+ modeling at the top level, and so on.
These truly integrate classic computer system design principles into the Agent architecture.
Understanding the underlying logic of AFK Comma and combining it with DeepSeek’s long-form content, we can already see the eventual direction of AI Agent development over the next three to five years.
Trend 1: The death of single‑point sessions, “one team, one brain” becomes the norm
In the future, your company and your life will each have only “one” super assistant (Router). All members and all apps (WeChat, Slack, Feishu) will feed their information into this central brain.
Trend 2: AI will shift from a “passive response” model to a “proactive advancement” model
Built on a dual‑layer architecture and a background silent Worker design, future AI will continuously monitor public group chats, GitHub, and email, becoming a proactive driver.
Trend 3: Cloud Brain, Everything Can Be an Agent
Because the environment and computation are separated, future execution nodes will sink into every employee’s computer, every phone, and even every smart home device. The AI’s brain resides in the cloud, using “hands” to control local resources.
Yesterday, DeepSeek published a long article telling us that for future superintelligent agents, a revolution in the underlying system is the real key.
The exciting part is that at this pivotal moment, a Chinese team in Silicon Valley is leading the way.
Edited by: David Aeneas
Get MirrorSpeed
Get MirrorSpeed