One-sentence summary
DeepSeek Harness published its Super-agent design on October 31: a two-tier, event-driven, async multi-channel architecture with compute-storage separation and an external Watchdog that handles session management and context sync, lifting humans to decisions and acceptance.
Why a Super-agent
As model capability grows, personal assistants are emerging. DeepSeek argues current mainstream agent designs cannot support multi-task and team collaboration at scale, where the ceiling is often human attention.
Key Design Ideas
Super-agent uses a two-tier architecture, event-driven async multi-channel interaction, compute-storage separation and an external Watchdog to take over tedious session management and context sync, moving humans up to higher-level decisions and acceptance.
Internal Instances Already Run
An internal Super-agent instance built on this design is already running, handling bug reproduction and fixes, code review responses and DeepSeek Harness's own iteration. Even when no one @s the AI after a teammate reports a bug in group chat, it can spot potential defects and try to reproduce and fix them in the background.
How It Differs from Single-Turn Agents
A normal single-turn agent relies on humans to issue step-by-step instructions and maintain context; Super-agent sinks session management, context sync and an exception watchdog to the system layer, leaving humans as supervisors and acceptors.
Sources
- Guancha (QQ News): November 1 hard-tech brief