DeepSeek Harness 官方团队于2024年10月31日发布《Super-agent 设计探索：迈向更自主的 Agent 系统》文章，正式提出面向工程实践的“Super-agent”架构设计 1。该架构旨在构建具备连续、高效、主动及持久特性的 Agent 系统，其核心采用 Router 与 Worker 的双层结构 1。其中，Router 层负责承接交互并维护上下文，而 Worker 层则通过隔离会话实现并发执行 1。为了支撑这一架构，团队引入了 Inline & Spill、链式召回（Chain of Recall）、投机执行以及异步 ReAct Loop 等关键机制，结合事件驱动与状态分离技术，以优化系统的并行处理能力 1。
在初步实验数据方面，Super-agent 在多任务处理效率上显示出优于传统单 Session 方案的潜力 1。数据显示，Super-agent 的任务完成时间中位数（p50）为12.4分钟，95分位值（p95）为15.7分钟；相比之下，中性提示下的单 Session 方案 p95 耗时高达29.5分钟 1。此外，压力测试结果表明，当 Trigger 速率从1条/分增加至40条/分时，单次请求耗时虽从2.5秒增至3.9秒，但单个决定分摊的推理时间却从6.4秒显著降至1.6秒 1。目前，DeepSeek 已提供名为 dsh-superagent-prototype 的原型，支持 BYOM（Bring Your Own Machine）模式，允许用户将本地设备接入云端 Agent 1。尽管内部实例已持续运行，但该架构仍处于非正式产品阶段，且团队指出长程记忆与安全隐私等问题仍待进一步解决 1。
The official team behind the DeepSeek Harness released an article on October 31, 2024, titled "Super-agent Design Exploration: Towards a More Autonomous Agent System," which outlines a new architectural framework designed for engineering practice. 1 This proposed "Super-agent" system aims to achieve greater autonomy, parallelism, and persistent online presence by utilizing a dual-layer structure composed of Router and Worker components. 1 The core concept focuses on creating agents that are continuous, efficient, proactive, and durable. 1
Architectural Components and Mechanisms
The architecture features a Router layer responsible for handling interactions and maintaining context, alongside a Worker layer that isolates sessions and executes tasks concurrently. 1 To support this structure, the design incorporates several key mechanisms, including Inline & Spill processing, Chain of Recall, speculative execution, and asynchronous ReAct loops. 1 Additionally, the system supports a Bring Your Own Machine (BYOM) mode, allowing users to connect local devices to cloud-based Agents. 1
Performance Results and Current Status
Initial experiments indicate that the Super-agent architecture outperforms traditional single-session solutions in multi-tasking efficiency. 1 Data shows that under neutral prompts, the Super-agent achieved a median task completion time of 12.4 minutes and a 95th percentile time of 15.7 minutes, compared to a 95th percentile time of 29.5 minutes for the single-session approach. 1 Stress tests revealed that as the trigger rate increased from one message per minute to forty messages per minute, the latency for a single request rose from 2.5 seconds to 3.9 seconds, while the inference time allocated per decision dropped from 6.4 seconds to 1.6 seconds. 1 An internal instance has been running continuously, providing a prototype known as dsh-superagent-prototype, though it is not yet a formal product. 1 The report also highlights remaining challenges, particularly regarding long-term memory capabilities and security and privacy concerns. 1
评论
还没有评论，欢迎留下第一条。