Has this Silicon Valley team implemented DeepSeek's newly unveiled next-generation Agent half a year ahead of schedule?

Recently, DeepSeek has once again made a huge splash in the industry.

Yesterday, DeepSeek unveiled the concept of its next-generation Agent Harness — Super-agent.

This long article points out clearly: The ultimate form of future Agents will completely break free from the constraints of the "chat window", and step into the era of "true Jarvis" with cross-session memory and multi-task concurrency.

The article sparked a huge stir across the industry as soon as it was released.

Many developers exclaimed: "This is exactly the next-generation AI application form we have been searching for painstakingly!"

However, while we are full of praise for DeepSeek's top-level architectural thinking, we have surprisingly discovered another fact —

As early as April this year, a Silicon Valley Chinese team named AFK AI (open source organization AFK-surf, product name Comma) had perfectly implemented almost all the core architectures described by DeepSeek today in engineering, and was months ahead of the rest of the world!

So, while industry giants are still releasing "concepts", how did this mysterious team turn the super architecture of "Single Session + Multi Worker" into reality?

And how will they bring superintelligence to everyone?

From now on, there are no redundant applications any more

Great Minds Think Alike: Declare War on the "Anti-human" Single Session!

Many people have found that they often run into frustrating situations when using AI: for example, when you are chatting with AI about code, you suddenly want it to help you reply to an email.

To avoid context pollution, you have to manually create a new conversation; when you finish replying to the email and switch back, you have to sort out the context all over again.

According to statistics, the human brain can only follow up 3 to 5 active sessions at the same time at most. The upper limit of the capability of traditional AI products is not the intelligence quotient of the model, but human attention.

In the latest article, DeepSeek pointed out sharply: The existing Agent design cannot support multi-tasking and team collaboration, and it is necessary to break the binding between Agent and a single manual session!

The system should be a continuous entity (Super-agent) that can maintain context across sessions, and advance multiple complex tasks simultaneously without interfering with each other.

Shockingly, this is exactly the same as the core direction internally determined by the AFK team on April 13 this year.

At that time, the AFK team put forward a resounding concept: "One team, one conversation" (Single session + Multi-worker).

They believe that after users gradually access AI via IM (Instant Messaging), they will inevitably mainly use the "Single Session" mode. No matter it is you, your colleagues, or any message in Feishu, Slack, WeChat, all information should be aggregated to a "main brain" with long-term memory.

DeepSeek summed up this pain point into a theory, while AFK had already submitted the code to GitHub months ago, and developed it into an extremely mature product called Comma.

For example, the automatic PR and cross-end task execution monitoring mentioned by DeepSeek have been running on AFK's system for more than half a year.

"Perfect Synchronization" of Core Architecture: Dual-layer Architecture and Environment Decoupling

If you compare DeepSeek's long article with AFK-surf's technical documents, you will get the illusion that the two parties "copied homework from each other".

But in fact, similar design constraints will inevitably lead to similar architecture choices. Only when people have truly touched the current ceiling of Agent implementation, will they evolve such a system by coincidence.

Let's take a look at their amazing technical similarities:

1. Dual-layer Architecture: Router (Main Brain) + Worker (Execution End)

To solve the contradiction of "needing to remember everything while handling many tasks at the same time", both parties abandoned the idea of "one large model doing all the work", and adopted the classic dual-layer architecture.

DeepSeek designed a resident Router layer and a Worker layer with isolated contexts.

Among them, the Router is responsible for taking overall control, receiving human interactions and assigning tasks, but it does not do the trivial and heavy work itself; specific tasks are sent to independent Workers, who work intensively in isolated contexts, and report back to the Router after finishing the work.

The solution of AFK Comma is completely consistent! AFK defines it as Lead Agent (Router) + Subagents (Workers).

Users or teams will always only talk to that long-term online Router. The Router is like a human supervisor, after receiving a task, it will create a new Session for the corresponding Worker.

Workers have their own environments, prompts and tools, multiple Workers run in parallel without interfering with each other, all complicated execution details remain on the Worker side, and only the results are aggregated to the Router.

The Router has a long-term main context called Single session, which represents "one person" or "one team", and can last for several months.

It remembers what happened before, and also knows what tasks are ongoing now. Workers are the hands it sends out to do the work, and it itself is the "master" who is always present and takes full responsibility. This solves the problem of "how humans and AI get along with each other in the long term".

This architecture brings a revolutionary product experience —

You no longer need to act as a group owner to create groups, nor do you need to set personas for AI every time. In Comma, you only need to speak, and the Router will take full charge of the rest of the routing and distribution work.

2. "Out-of-body Experience": Separation of Computing and Execution Environment (BYOM)

Ordinary AI can only execute tasks in the virtual sandbox on the cloud, once it involves code on the corporate intranet or your local private files, the AI will become completely blind.

DeepSeek proposes that computing, status and environment must be separated.

The "brain" of the Agent is on the cloud, but its "hands" can extend to the user's local machine. By deploying a lightweight stateless Daemon on the end side, the cloud Agent can read the local environment on demand, which is called BYOM (Bring Your Own Machine).

The implementation of AFK Comma is one step ahead again — Comma has turned this concept into a powerful enterprise-level feature!

By @Comma bot in the enterprise IM, the Agent can request the employee's terminal to execute tasks. The employee's personal Personal agent can take over the personal computer as a task execution node. At the same time, when someone in the enterprise @Comma bot, it will trigger the scheduling and communication mechanism between Agents, inherit human permissions and roles, and let AI collaborate and work for people without sacrificing information security.

At the same time, this also means that the network composed of Agents can gain insight into the subtle business details of the enterprise and achieve continuous learning.

Meanwhile, the system has a built-in security protection mechanism for information flow control (ifc), which manages the access and transmission of sensitive data to ensure information security.

This is equivalent to the fact that DeepSeek has just demonstrated the feasibility of "brain-hand separation", while AFK Comma has already sold this set of "cloud brain + local hand" system to enterprises for use!

3. Event-driven and Multi-end Concurrency: Real "Full Duplex"

Human work is not "one question, one answer", but asynchronous.

To this end, DeepSeek proposes to introduce event-driven and full-duplex interaction (asynchronous ReAct Loop). The model runs tasks in the background, and the front end can still reply to your new messages in seconds.

The client of AFK Comma has directly connected to Telegram, Signal, WeChat, iMessage, Feishu and Slack!

zanweiguo said straightforwardly: From now on, there is no need to open any other apps.

It is a real "workstation". The tasks you assign in Feishu will be sent by the Router to Workers for execution in the background; when colleagues ask about the progress in Slack, the Router can already give a synchronized reply.

Even, all tasks generated from these major platforms are uniformly managed in the Comma client; in turn, Comma can also create and manage tasks on third-party platforms.

This is the real Jarvis. It is ubiquitous, with absolutely consistent multi-end memory, not bound by any single App, but can control all information flows.

Multiple Demo Demonstrations

Below are more amazing demos of Comma.

It can help you manage Slack.

It can even fix itself:

It can also easily handle the ancient style 3D mahjong game.

You can even play 3D mahjong with it, which is absolutely amazing!

It can help you track flight tickets.