DeepSeek Outlines a Super-agent Architecture for Autonomous Collaboration
DeepSeek Harness’s official team published a Super-agent architecture proposal centered on Router and Worker layers, event-driven execution, persistent state, and an independent Watchdog. The design also includes asynchronous tool calls, speculative execution, and a dsh-superagent-prototype implementation for validating the approach, rather than a formal DeepSeek product. In vendor-reported tests using synthetic workloads, Super-agent recorded a p95 task-completion time of 15.7 minutes, versus 29.5 minutes for a neutral-prompt single-session setup.
The proposal offers a concrete engineering blueprint for agents that coordinate multiple concurrent tasks across channels such as IM and GitHub, but its performance evidence is vendor-reported and does not measure individual task-execution ability.
IT之家 10 月 31 日消息，DeepSeek Harness 官方团队今日发布《Super-agent 设计探索：迈向更自主的 Agent 系统》一文，介绍了其面向工程实践的 Super-agent 架构设计思路。
随着大模型执行单个任务时越来不需要人的干预，限制生产力进一步提升的因素正逐渐转向人的注意力，Agent 系统需要从被动响应指令的单任务工具，向持续在线、能够并行处理多项任务的自主协作者演进。
DeepSeek 表示，现有主流 Agent 设计仍以单一 Session（会话）为基本单位，不同 Session 之间相互独立。用户并行推进多项工作时，需要手动切换并维护各个 Session 的上下文。
根据 DeepSeek 团队观察，一个人通常只能同时跟进 3 至 5 个活跃 Session。即使通过 CLI、MCP 或外部连接器扩展本地 Agent，也仍可能面临单会话同步阻塞、多任务上下文混乱，以及本地设备离线后任务停止等问题。
#DeepSeek#DeepSeek Harness#Super-agent#dsh-superagent-prototype#Router#Worker#Watchdog