DeepSeek unveils new architecture: one AI orchestrating multiple AI agents

Chinese AI startup DeepSeek has released a new AI agent architecture that lets one AI orchestrate multiple AIs to complete tasks, simulating a virtual team made up of AIs.

According to reports from Chao News, IT Home and other outlets, the DeepSeek Harness team published an article on its official WeChat account on the evening of Saturday, October 31, unveiling an AI agent architecture called Super-agent.

Under the concept, AI is no longer confined to a single chat window; instead, it stays continuously online, runs multiple tasks in parallel, and proactively responds to changes in its environment.

The prototype has been open-sourced, and an internal instance is already running in DeepSeek's day-to-day research and development.

The DeepSeek team noted that while AI can now complete an increasing number of tasks, humans' attention for managing those tasks has not grown in step—a person can typically keep up with only three to five active sessions at once.

What Super-agent aims to do is move session management and context synchronization from people to the system level: humans set goals and sign off on results, while scheduling and execution are left to AI.

Specifically, Super-agent consists of two layers, Router and Worker, resembling a virtual team made up of AIs.

The Router is a continuously running global coordination node that receives messages from platforms such as group chats and email in a unified way and assigns tasks; each Worker has its own independent context and is responsible for problem reproduction, code review, and feature development, and can call on Subagents when necessary.

What users face is always the same AI assistant, while behind the system multiple agents may be working simultaneously.

In addition, Super-agent is proactive: it can listen silently in public group chats, proactively identify problems, and submit fix proposals.

However, the team noted at the end of the article that they are not proposing an entirely new theory—they are simply applying classic computer systems ideas such as task scheduling, event-driven design, and the separation of storage and compute to agent architecture—and that the experiment did not evaluate real coding quality.

But as AI increasingly completes individual tasks without human intervention, how many AIs a human can manage at once is becoming a new question.