ChainThink reports that on October 31, the DeepSeek Harness team released a technical article titled "Super-agent Design Exploration: Toward More Autonomous Agent Systems," introducing an agent architecture designed for multi-tasking and team collaboration.
The team noted that existing agents typically operate on a single-session basis, with each session independent of others, limiting the scale of multi-task collaboration.
The new architecture manages sessions and context synchronization through a two-layer design, event-driven processing, asynchronous multi-channel interaction, compute-storage separation, and an external watchdog. DeepSeek refers to this form as a Super-agent.
Internal instances built on this approach have been running continuously, participating in issue reproduction and resolution, code review responses, and the iteration of DeepSeek Harness itself.


