Chinese artificial intelligence (AI) start-up DeepSeek has unveiled a new AI agent architecture that allows one AI to coordinate and direct multiple agents to complete tasks, creating a virtual team of AI agents.
According to reports by Chao News, IT Home and Huanqiu.com, the DeepSeek Harness team published a technical article on its WeChat official account on Saturday (31 October) evening, unveiling a new AI agent architecture called Super-agent.
Under this vision, AI is no longer confined to a single chat window; instead, it can stay continuously online, advance multiple tasks simultaneously and proactively respond to changes in its environment. Relevant prototypes have been open-sourced, and internal versions are already running as part of DeepSeek’s day-to-day research and development.
The DeepSeek Harness team noted that while AI can handle an increasing range of tasks, humans’ capacity to manage them has not grown in tandem. A person can typically keep track of only three to five active conversations at once.
Super-agent aims to shift conversation management and context synchronisation from humans to the system. AI handles coordination and execution, while people focus on setting goals and approving results.
Super-agent consists of two layers, Router and Worker, resembling a virtual team of AI agents. Router is a continuously running central coordination node that receives messages from multiple channels, such as group chats and email, and assigns tasks. Workers each maintain their own independent context and work in parallel on specific tasks such as issue reproduction, code review and feature development.
This means users always interact with the same AI assistant, even though multiple agents may be working simultaneously behind the scenes.
The DeepSeek Harness team said in the article that it had not proposed an entirely new theory, but had applied established engineering principles from computer systems to building agent systems. While implementation methods would evolve with technological advances, the underlying system design principles would remain useful as a reference over the longer term. “As model capabilities and underlying system architectures evolve together, this more autonomous way of working could become the norm for human-AI collaboration.”
