DeepSeek Harness published Exploring Super-agent Design—an architecture exploration plus a research prototype, not a finished commercial product. The core claim: single-task limits track the model; multi-task limits track human attention. Super-agent pushes people to set goals and accept work; a long-lived Router dispatches while isolated Workers run tasks. On the open-source side you can inspect lobehub/lobehub (~84k GitHub stars as of check), lobehub/acceptance for evidence-based handoff, and dsh-superagent-prototype. A minimal usable pattern is manager + writer/reviewer: one outward voice, people only Yes/No on evidence. Unverified speed-up multiples and second-hand overnight demos stay out of the main claims.
Sources: https://teapawt.org/DeepSeekHarness/article/2116857257245987236 · https://github.com/deepseek-ai/dsh-superagent-prototype · https://github.com/lobehub/lobehub · https://github.com/lobehub/acceptance · https://lobehub.com/ · verified 2026-11-07 Asia/Shanghai.