DeepSeek Harness team outlines “Super-agent” design for more autonomous AI systems

The DeepSeek Harness team said in a technical article on its official WeChat account that it has outlined a “Super-agent” design aimed at building more autonomous agent systems. DeepSeek said gains in model capabilities have spurred the emergence of personal agents, but existing mainstream agent designs struggle to support multi-tasking and team collaboration at scale, with multi-task limits tied to human attention. The team said its approach uses a two-tier architecture, event-driven operation, asynchronous multi-channel interaction, separation of storage and computation and an external watchdog to overcome these limits. The system takes over cumbersome session management and context synchronization at the underlying layer, allowing humans to move from execution details to higher-level decision-making and acceptance. DeepSeek said internal Super-agent instances built on the approach are already running continuously, participating in issue reproduction and fixes, code review responses and iteration of DeepSeek Harness itself. When team members report software problems in group chats, the agent can identify potential defects even without being mentioned and try to reproduce the issue and find fixes in the background.