Oct. 31 — DeepSeek’s Harness team published a technical paper, Super-agent
design exploration: Toward more autonomous agent systems, outlining an
architecture to scale agents beyond single-session limits. The paper says rising
model capability has spurred personal assistant agents but current designs
isolate sessions and cap multi-tasking at model ability and human attention.
DeepSeek proposes a two-layer architecture with event-driven, asynchronous
multi-channel interaction, compute–storage separation and an external watchdog
to centralize session management and context sync, shifting humans to
higher-level decision and acceptance tasks. The firm says internal Super-agent
instances are already running for bug reproduction and repair, code-review
responses and iterative development of DeepSeek Harness.
