On the evening of October 31, the DeepSeek Harness team published a technical article titled "Exploring Super-agent Design: Towards More Autonomous Agent Systems." DeepSeek stated that the enhancement of model capabilities has given rise to new application forms, with personal assistants (Personal Agents) recently emerging in large numbers. However, existing mainstream Agent designs struggle to support this model at the scale of multitasking and team collaboration, as the working mode of Agents is still based on a single session, with different sessions operating independently. Under this design, the limit of single-task performance depends on model capabilities, while the limit of multitasking depends on human attention. By employing a dual-layer architecture, event-driven design, asynchronous multi-channel interaction, separation of storage and computation, and external watchdog concepts, the DeepSeek team has broken through these limitations. The system takes over cumbersome session management and context synchronization at the underlying level, allowing humans to shift from specific execution details to higher-level decision-making and acceptance. DeepSeek refers to this form of Agent and its system architecture as Super-agent. The article mentioned that currently, internal Super-agent instances built on this concept are continuously running, participating in daily problem reproduction and resolution, code review responses, and the iterative development of DeepSeek Harness itself.
All Comments