DeepSeek presents Super-agent two-layer Router and Worker architecture for autonomous agent systems in engineering practice
DeepSeek Harness's official team published an article on October 31 titled an exploration of Super-agent design toward more autonomous agent systems, outlining architecture ideas aimed at engineering practice.
DeepSeek Harness's official team published an article on October 31 titled an exploration of Super-agent design toward more autonomous agent systems, outlining architecture ideas aimed at engineering practice. As models increasingly execute single tasks without human intervention, the factor limiting productivity is shifting to human attention, so agent systems need to evolve from single-task tools that passively answer instructions into always-on collaborators that handle several tasks in parallel. DeepSeek noted mainstream agent design still uses a single session as its basic unit with sessions independent of each other, and a person typically keeps only three to five active sessions in flight. The team calls the new form and its architecture Super-agent and summarizes four traits: continuity, efficiency, proactivity and persistence. The overall architecture has five parts, a two-layer design, event-driven operation, human-facing interaction, separation of compute from state and environment, and a Watchdog. The two-layer design splits the system into a Router layer that takes in human interaction, maintains global context and dispatches tasks, and Worker layers running concrete work concurrently in isolated sessions, reporting status and results to the Router rather than replying to users directly. Group messages, email and calendar reminders, code review comments and CI results serve as Triggers, filtered first by program rules and then judged by a model for whether action is needed. The team also described Inline and Spill and a Chain of Recall mechanism, plus speculative execution and an asynchronous ReAct Loop to cut waiting time.