DeepSeek Publishes Super-agent Architecture and Open-Sources Research Prototype
On October 31, 2026, the DeepSeek Harness team published "Exploring Super-agent Design: Towards a More Autonomous Agent System," describing a Router/Worker two-layer architecture with event-driven triggers, asynchronous multi-channel interaction, separated compute and state, and an external Watchdog session. Parts of the design are implemented in the MIT-licensed dsh-superagent-prototype on GitHub, which the team says is only a prototype for validating the technical approach and not an official DeepSeek product; the public version targets a single GitHub repository with optional Feishu integration. DeepSeek's own synthetic-load tests, using about 300 messages with roughly 20 work requests and five repeats per scheme, reported task-completion p50/p95 of 12.4 and 15.7 minutes versus 29.5 minutes p95 for a single-session baseline, and an internal Super-agent instance has been running on issue reproduction, code-review responses and DeepSeek Harness iteration.
The Super-agent design moves agent orchestration from single-session chat tools into a persistent, event-driven system layer that handles parallel task scheduling and context management, a direction Bloomberg and other outlets linked to DeepSeek's push to compete with Anthropic's Claude Code.
11月1日消息，DeepSeek Harness 官方团队发布《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越不需要人的干预，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。
他们将这种新型 Agent 形态及其系统架构称为 Super-agent，并归纳出四项基础特征：连续、高效、主动和持久。
#DeepSeek#DeepSeek Harness#dsh-superagent-prototype#Anthropic#Claude Code#Feishu