DeepSeek Proposes Super-agent Architecture Design, Exploring More Autonomous Agent Systems

October 31 — According to IT之家 (ITHome), the official DeepSeek Harness team released "Super-agent Design Exploration: Toward a More Autonomous Agent System," introducing its engineering-practice-oriented Super-agent architecture. DeepSeek said that existing Agents operate on a single Session basis, so users must switch manually when handling multiple tasks in parallel; typically only 3 to 5 active Sessions can be followed at once, and users also face issues such as single-session blocking, context confusion, and tasks stopping once the local machine goes offline.

DeepSeek summarizes the characteristics of Super-agent as continuous, efficient, proactive and persistent, and its overall architecture consists of five parts: the two-layer Router and Worker structure, event-driven design, human-oriented interaction, separation of compute from state and environment, and Watchdog. The Router handles human interaction, maintains global context and dispatches tasks; Workers execute concurrently in isolated sessions and do not reply to users directly, but instead report their status and results to the Router. The system also introduces mechanisms including Trigger event filtering, Inline & Spill, chained recall, speculative execution, asynchronous ReAct Loop and Watchdog.