AI资讯大全AIPPXP.CN搜索 ↗
AI 快讯 · 自动更新

DeepSeek proposes Super-agent architecture design: Router and Worker two-tier architecture, explore more autonomous agent systemDeepSeek 提出 Super-agent 架构设计:Router 与 Worker 双层架构,探索更自主的 Agent 系统

IT House October 31st news, DeepSeek Harness official team today released "Super-agent design exploration: towards a more autonomous Agent system" article, introducing its engineering practice-oriented Super-age...

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-age…

2026年10月31日 发布 · 1 分钟阅读 · 免费
DeepSeek 提出 Super-agent 架构设计:Router 与 Worker 双层架构,探索更自主的 Agent 系统

Key Points要点速览

IT House October 31st news, DeepSeek Harness official team today released "Super-agent design exploration: towards a more autonomous Agent system" article, introducing its engineering practice-oriented Super-agent architecture design ideas. As the large model performs a single task without human intervention, the factors that limit further productivity improvement are gradually shifting to human attention, and the Agent system needs to evolve from a single-task tool that responds to instructions passively to an autonomous collaborator who is continuously online and can handle multiple tasks in parallel. DeepSeek said that the existing mainstream Agent design still uses a single session as the basic unit, and different sessions are independent of each other. When a user advances multiple tasks in parallel, the context of each session needs to be manually switched and maintained. According to the DeepSeek team, a person can typically only follow 3 to 5 active sessions at a time. Even if you extend the local agent through the CLI, MCP, or external connectors, you may still face problems such as single-session synchronization blocking, multitasking context confusion, and task stopping after the local device goes offline. They call this new form of Agent and its system architecture Super-agent, and summarize four basic characteristics: continuous, efficient, proactive, and durable. First of all, continuous means that users can interact with the same Agent in different channels such as instant messaging and GitHub comments; efficient means that they can handle multiple complex tasks in parallel; active means that the system can continuously observe environmental events and promote things worth dealing with on its own; durable requires that the Agent is always online and can recover from abnormal conditions during long-term operation. The ▲ same task continuously promotes ▲ continuous attention to public group chat between IM group chat and GitHub Pull Request review. In the case of no one @, the overall architecture proposed by Pull Request DeepSeek is mainly composed of a two-tier architecture, event-driven, human-oriented interaction, separation of computing and state and environment, and Watchdog. Among them, the two-tier architecture divides the system into a Router layer and a Worker layer: the Router is responsible for accepting human interactions, maintaining the global context, and assigning tasks, while the Workers perform specific tasks concurrently in isolated sessions. Instead of responding directly to the user, the worker reports the status and results to Router, who communicates with the user. ▲ Two-layer architecture: The Router layer is responsible for routing and external interaction. Multiple Worker Sessions are isolated and concurrently execute tasks. In the event-driven mechanism, external information such as group chat messages, email and schedule reminders, code review comments, and CI operation results are called triggers. The system first filters events through procedural rules, and then leaves it to the large model to determine whether action is needed. For event streams that have been assigned tasks, the worker can subscribe to the corresponding Thread, and subsequent related events are sent directly to the worker by the program, eliminating the need to go through the Router every time, thereby reducing the routing load. The DeepSeek team also introduced the Inline & Spill and Chain of Recall mechanisms. The former places information that may be used in the near future directly into the context window of the model, and retains the content beyond the budget or low-frequency use outside the window for the model to read on demand; The latter allows the model to gradually retrieve historical task information along the association ID through the reference relationship between objects such as Trigger, Thread, and Worker Session. ▲ Super-agent overall system architecture In order to reduce human waiting time, Super-agent also introduced a speculative execution mechanism that can work in advance in the background according to environmental signals. For example, after receiving the meeting invitation, check the schedule conflict, collect background information, and prepare a draft response; in the R&D collaboration scenario, you can try to reproduce the fault and find a fix during the group chat to discuss the problem, and deliver the result to the user only after confirming the problem and finding a simple fix. In terms of human-computer interaction, the article proposes asynchronous ReAct Loop, changing the time-consuming tool call to asynchronous start () and stop (). After starting the background task, the Agent can return to the Started state without waiting for the task to complete. The background result then enters the session as a new event. This way, the foreground conversation can continue to receive new messages from the user. However, the DS team specifically stated that this asynchronous task mechanism has been implemented in DeepSeek Harness, but has not been turned on by default. ▲ Two-level filtering: First, the program layer policy filters out most of the irrelevant messages, and then the semantic layer policy judges that for long-term online operation, DeepSeek decouples the calculation, session state and execution environment of Agent Loop: the calculation and persistence state are placed in the cloud, and the execution environment can be kept on the user's local device, and only the lightweight daemon responsible for executing commands and returning results is deployed. The design supports BYOM (Bring Your Own Machine), that is, the development machine or personal computer that users use daily to connect to the cloud Agent, and also supports connecting to cloud-hosted machines or on-demand computing environments. In addition, the system sets up a separate Watchdog Session outside the main architecture, forming an external control loop. Watchdog not only monitors infrastructure issues such as interface errors and state recovery failures, but also judges whether the worker has made substantial progress and whether the execution direction deviates from the target through the model, and intervenes to repair when abnormalities are found. ▲ Thread example: A Thread DeepSeek team composed of consecutive Triggers under the same IM topic or the same GitHub Pull Request also conducted preliminary experiments to compare the Super-agent, the Super-agent without the event pass-through mechanism, and the single Session plus Subagent scheme under the two prompt configurations. The experiment consisted of about 300 messages at a time, of which about 20 were work requests, each of which was repeated 5 times. The results showed that the task completion time p50 and p95 of the Super-agent were 12.4 minutes and 15.7 minutes, respectively; the single session scheme p95 under the neutral prompt was 29.5 minutes. DeepSeek team pointed out that these experiments are based on synthetic load and simplified environment, mainly to evaluate the system scheduling and collaboration ability, and do not measure the actual ability to perform individual tasks. The cumulative arrival curve of the ▲ work request and the cumulative completion curve of the four schemes (average of 5 experimental runs; the message interval and working time are scaled by 1/4, and the model call time is the real value) In the Router layer stress test, when the arrival rate of the Trigger is increased from about 1 per minute to about 40 per minute, the average batch size increases from 1.0 to 2.8, the time required for a single request increases from 2.5 seconds to 3.9 seconds, and the inference time for a single decision to allocate decreases from 6.4 seconds to 1.6 seconds. Based on this, the team believes that in the event of an outbreak, Router can improve effective throughput by merging the backlog of events. ▲ Each scenario runs 5 times repeatedly; task completion time refers to the time taken from the work request to the result message. At the same time, the article pointed out that there are still problems to be solved in this architecture, including how the Router layer scales with the scale of tasks, long-term memory across months and even longer, the way of human-computer collaboration after leaving the Session, A2A communication between different Agent instances, and security and privacy protection. Especially after the cloud Agent connects to the local device, prompt word injection, end-side credential leakage, and cross-device permission risks need to be further solved. ▲ The DeepSeek team said that the internal Super-agent instances built based on the above ideas and DeepSeek Harness have been continuously running, participating in problem reproduction and repair, code review response, and DeepSeek Harness itself iteration. In addition, they provide a prototype implementation called dsh-superagent-prototype in the article, but make it clear that the plugin is used to validate the technical route and is not an official DeepSeek product. Reference: GitHub repository for dsh-superagent: https://github. com/deepseek-ai/dsh-superagent-prototype

中文

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。 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。 首先,连续意味着用户可以在即时通讯、GitHub 评论等不同渠道与同一个 Agent 交互;高效意味着能够并行处理多个复杂任务;主动意味着系统可以持续观察环境事件并自行推进值得处理的事项;持久则要求 Agent 始终在线,并能在长期运行中从异常状态恢复。 ▲ 同一个任务在 IM 群聊与 GitHub Pull Request 评审之间连续推进 ▲ 持续关注公共群聊,在无人 @ 的情况下主动识别缺陷并提交修复 Pull Request DeepSeek 提出的整体架构主要由双层架构、事件驱动、面向人的交互、计算与状态及环境分离、Watchdog 五部分组成。 其中,双层架构将系统划分为 Router 层和 Worker 层:Router 负责承接人类交互、维护全局上下文并分派任务,Worker 则在相互隔离的会话中并发执行具体工作。Worker 不直接向用户回复,而是将状态和结果汇报给 Router,由后者统一对外沟通。 ▲ 双层架构:Router 层统一负责路由与对外交互,多个 Worker Session 隔离并发执行任务 在事件驱动机制中,群聊消息、邮件与日程提醒、代码审查评论、CI 运行结果等外部信息被称为 Trigger。系统先通过程序规则过滤事件,再交由大模型判断是否需要采取行动。对于已经分配任务的事件流,Worker 可以订阅对应 Thread,后续相关事件由程序直接发送给 Worker,无需每次都经过 Router,从而降低路由负载。 DeepSeek 团队还介绍了 Inline & Spill 和链式召回(Chain of Recall)机制。前者将近期可能用到的信息直接放入模型上下文窗口,将超出预算或低频使用的内容保留在窗口之外,供模型按需读取;后者则通过 Trigger、Thread 和 Worker Session 等对象之间的引用关系,让模型沿关联 ID 逐步找回历史任务信息。 ▲ Super-agent 整体系统架构 为减少人类等待时间,Super-agent 还引入投机执行机制,可以根据环境信号在后台提前开展工作。例如,收到会议邀约后先检查日程冲突、搜集背景资料并准备回复草稿;在研发协作场景中,则可以在群聊讨论问题时尝试复现故障、寻找修复方案,只有确认问题并找到简单修复方式后才向用户交付结果。 在人机交互方面,文章提出异步 ReAct Loop,将耗时工具调用改为异步的 start () 和 stop ()。Agent 启动后台任务后即可返回 Started 状态,不必等待任务完成,后台结果随后作为新事件进入会话。 这样,前台对话可以继续接收用户的新消息。不过,DS 团队特别说明,这项异步任务机制已经在 DeepSeek Harness 中实现,但尚未默认开启。 ▲ 两级过滤:先由程序层策略筛除大部分无关消息,再交由语义层策略判断 针对长期在线运行,DeepSeek 将 Agent Loop 的计算、会话状态和执行环境进行解耦:计算与持久化状态放在云端,执行环境则可以保留在用户本地设备,仅部署负责执行命令和返回结果的轻量 daemon。该设计支持 BYOM(Bring Your Own Machine),即将用户日常使用的开发机或个人电脑接入云端 Agent,也支持连接云端托管机器或按需创建的计算环境。 此外,系统在主架构之外设置独立的 Watchdog Session,形成外部控制回路。Watchdog 不仅监控接口报错、状态恢复失败等基础设施问题,还会通过模型判断 Worker 是否取得实质进展、执行方向是否偏离目标,并在发现异常时介入修复。 ▲ Thread 示例:同一个 IM 话题或同一个 GitHub Pull Request 下的连续 Trigger 构成一条 Thread DeepSeek 团队还进行了初步实验,对比 Super-agent、去掉事件直通机制的 Super-agent,以及两种提示词配置下的单 Session 加 Subagent 方案。 实验每次包含约 300 条消息,其中约 20 条为工作请求,每种方案重复运行 5 次。结果显示,Super-agent 的任务完成时间 p50 和 p95 分别为 12.4 分钟和 15.7 分钟;中性提示下的单 Session 方案 p95 为 29.5 分钟。 DeepSeek 团队指出,这些实验基于合成负载和简化环境,主要评估系统调度与协作能力,并不衡量单个任务的实际执行能力。 ▲ 工作请求的累计到达曲线与四种方案的累计完成曲线(5 次实验运行均值;消息间隔与工作时长按 1/4 缩放,模型调用耗时为真实值) 在 Router 层压力测试中,Trigger 到达速率从每分钟约 1 条提升至约 40 条时,平均批处理大小从 1.0 增至 2.8,单次请求耗时从 2.5 秒增至 3.9 秒,而单个决定分摊的推理时间从 6.4 秒降至 1.6 秒。团队据此认为,事件突发时,Router 可以通过合并积压事件提高有效吞吐。 ▲ 每种方案重复运行 5 次;任务完成时间指从工作请求到结果消息的耗时。 文章同时指出,这套架构仍存在待解决的问题,包括 Router 层如何随任务规模扩展、跨月乃至更长时间跨度的长程记忆、脱离 Session 后的人机协作方式、不同 Agent 实例之间的 A2A 通信,以及安全与隐私保护。尤其是云端 Agent 接入本地设备后,提示词注入、端侧凭据泄露和跨设备权限风险都需要进一步解决。 ▲ 不同 Trigger 到达速率下 Router Session 的 batch 大小、单次请求耗时与单个决定的分摊推理时间 DeepSeek 团队表示,基于上述思路和 DeepSeek Harness 构建的内部 Super-agent 实例已经持续运行,参与问题复现与修复、代码审查响应以及 DeepSeek Harness 自身迭代。 此外,他们在文章中还提供了名为 dsh-superagent-prototype 的原型实现,但明确指出,该插件用于验证技术路线,并非 DeepSeek 正式产品。 参考资料: dsh-superagent 的 GitHub 仓库: https://github.com/deepseek-ai/dsh-superagent-prototype

读完接着看 · Keep reading

想马上用起来?去「AI工具」栏挑一个直接下载,或在「AI教程」里跟着图文步骤做一遍。

去 AI工具 → 看 AI教程 →
0阅读0 条评论

阅读与点赞数据保存在你的浏览器本地,欢迎留下你的想法。

评论 文明发言,让讨论更有价值

正能量公益广告今日正能量学一点,用一点;今天种下的种子,会长成明天的能力。去免费下载专区 →广告