DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents


AI free token/AI免费积分
感谢IT之家网友 xxy171070、软媒新友2547452、Wllen、Tokishio、阿狗 的线索投递!

IT之家 10 月 31 日消息,DeepSeek Harness 官方团队今日发布《Super-agent 外观探索:迈向更自主的 Agent 平台》一文,介绍了其面向工程实践的 Super-agent 架构外观思路。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

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高的因素正逐步转向人的注意力,Agent 平台需要从被动响应指令的单任务工具,向持续在线、可以并行处理多项任务的自主协作者演进。

DeepSeek 表示,现有主流 Agent 外观仍以单一 Session(会话)为基本单位,不同 Session 之间相互独立。使用者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。

依据 DeepSeek 团队观察,一个人通常只能同时跟进 3 至 5 个活跃 Session。即使借助 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难题。

他们将这种新型 Agent 形态及其平台架构称为 Super-agent,并归纳出四项基础特征:连续、高质、主动和持久。

首先,连续意味着使用者可以在即时通讯、GitHub 评论等不同渠道与同一个 Agent 交互;高质意味着可以并行处理多个复杂任务;主动意味着平台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。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 同一个任务在 IM 群聊与 GitHub Pull Request 评审之间连续推进

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 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

其中,双层架构将平台划分为 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 不直接向使用者回复,而是将状态和结果汇报给 Router,由后者统一对外沟通。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 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。平台先借助程序规则过滤事件,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;后者则借助 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。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ Super-agent 总体平台架构

为减少人类等待时间,Super-agent 还引入投机执行机制,可以依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;在研发协作情形中,则可以在群聊讨论难题时尝试复现故障、寻找修复思路,只有确认难题并找到简单修复方式后才向使用者交付结果。

在人机交互维度,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使用者的新消息。不过,DS 团队特别说明,这项异步任务机制已经在 DeepSeek Harness 中达成,但尚未默认启动。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 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开展解耦:计算与持久化状态放在云端,执行环境则可以保留在使用者本地设备,仅部署负责执行命令和返回结果的轻量 daemon。该外观兼容 BYOM(Bring Your Own Machine),即将使用者日常使用的开发机或个人电脑接入云端 Agent,也兼容连接云端托管机器或按需创建的计算环境。

另外,平台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难题,还会借助模型判断 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。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 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 团队还开展了初步实验,比较 Super-agent、去掉事件直通机制的 Super-agent,以及两种提示词规格下的单 Session 加 Subagent 思路。

实验每次包含约 300 条消息,其中约 20 条为工作请求,每种思路重复运行 5 次。结果表明,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 思路 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,主要评估平台调度与协作能力,并不衡量单个任务的实际执行能力。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 工作请求的累计到达曲线与四种思路的累计完成曲线(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 条提高至约 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 可以借助合并积压事件提高有效吞吐。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 每种思路重复运行 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,这套架构仍存在待应对的难题,涵盖 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应对。

DeepSeek 提出 Super-agent 架构:Router 与 Worker 双层设计,探索更自主的 Agent 系统

▲ 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,参与难题复现与修复、代码审查响应以及 DeepSeek Harness 自身迭代。

另外,他们在文章中还给出了名为 dsh-superagent-prototype 的原型达成,但明确指出,该插件用于验证工艺路线,并非 DeepSeek 正式机型。

参考资料:

本文来源:https://www.ithome.com/1/020/615.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Thanks to IT之家 netizens xxy171070, ruanmei-friend2547452, Wllen, Tokishio, and 阿狗 for sending in the leads!

According to IT之家 on October 31, the official DeepSeek Harness team published an article today titled "Super-agent Appearance Exploration: Towards a More Autonomous Agent Platform," introducing their Super-agent architecture concept aimed at engineering practices.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

As large models increasingly require less human intervention when executing individual tasks, the limiting factor for further productivity improvements is gradually shifting towards human attention. Agent platforms need to evolve from single-task tools that passively respond to instructions into continuously online, autonomous collaborators capable of processing multiple tasks in parallel.

DeepSeek stated that the current mainstream Agent appearance still uses a single Session as the basic unit, with different Sessions being mutually independent. When users work on multiple tasks in parallel, they need to manually switch and maintain the context of each Session.

Based on observations by the DeepSeek team, an individual can typically only keep track of 3 to 5 active Sessions simultaneously. Even when using CLI, MCP, or external connectors to extend local Agents, challenges such as single-session synchronization blocking, multi-task context confusion, and task halting after local devices go offline may still arise.

They refer to this new form of Agent and its platform architecture as Super-agent, summarizing four fundamental characteristics: continuous, high-quality, proactive, and persistent.

First, continuous means users can interact with the same Agent across different channels like instant messaging and GitHub comments; high-quality means it can handle multiple complex tasks in parallel; proactive means the platform can continuously observe environmental events and autonomously advance matters worth handling; persistence requires the Agent to remain always online and capable of recovering from abnormal states during long-term operation.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

Specifically, the dual-layer architecture divides the platform into a Router layer and a Worker layer: the Router is responsible for handling human interactions, maintaining global context, and dispatching tasks, while Workers execute specific work concurrently in isolated Sessions. Workers do not reply directly to users; instead, they report status and results to the Router, which handles all external communication uniformly.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

In the event-driven mechanism, external information such as group chat messages, emails, calendar reminders, code review comments, and CI run results are referred to as Triggers. The platform first filters events using program rules and then uses large models to determine if action is required. For event streams where tasks have been assigned, Workers can subscribe to corresponding Threads, and subsequent relevant events are sent directly to the Workers by the program without passing through the Router every time, thereby reducing routing load.

The DeepSeek team also introduced the Inline & Spill and Chain of Recall mechanisms. The former places recently potentially useful information directly into the model's context window, while keeping content that exceeds the budget or is low-frequency usage outside the window for the model to read on demand; the latter utilizes reference relationships between objects like Triggers, Threads, and Worker Sessions to allow the model to progressively retrieve historical task information along associated IDs.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

To reduce human waiting time, Super-agent has also introduced a speculative execution mechanism, which can start work in the background ahead of time based on environmental signals. For example, after receiving a meeting invitation, it can first check for schedule conflicts, collect background materials, and prepare a reply draft; in R&D collaboration scenarios, it can attempt to reproduce faults and search for fix ideas while discussing difficulties in group chats, delivering results to users only after confirming the issue and finding a simple fix.

In the dimension of human-computer interaction, the article proposes an asynchronous ReAct Loop, changing time-consuming tool calls to asynchronous start () and stop (). After launching a background task, the Agent can return a Started status without waiting for the task to complete, and the background result later enters the Session as a new event.

This allows the front-end dialogue to continue receiving new messages from users. However, the DS team specifically noted that this asynchronous task mechanism has already been achieved in DeepSeek Harness but is not enabled by default yet.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

Regarding long-term online operation, DeepSeek decouples the computation, Session state, and execution environment of the Agent Loop: computation and persistent state are placed in the cloud, while the execution environment can remain on the user's local device, deploying only a lightweight daemon responsible for executing commands and returning results. This appearance supports BYOM (Bring Your Own Machine), connecting users' daily development machines or personal computers to the cloud Agent, and also supports connecting to cloud-hosted machines or on-demand computing environments.

Additionally, the platform sets up an independent Watchdog Session outside the main architecture, forming an external control loop. The Watchdog not only monitors infrastructure issues such as interface errors and state recovery failures, but also uses models to determine whether Workers are making substantial progress and if their execution direction is deviating from the goal, intervening to fix problems when anomalies are detected.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

The DeepSeek team also conducted preliminary experiments comparing Super-agent, Super-agent without the direct event routing mechanism, and a single Session plus Subagent approach under two prompt specifications.

Each experiment contained approximately 300 messages, of which about 20 were work requests, and each approach was run 5 times. The results showed that the task completion times for Super-agent had p50 and p95 of 12.4 minutes and 15.7 minutes respectively; the single Session approach under neutral prompts had a p95 of 29.5 minutes.

The DeepSeek team pointed out that these experiments were based on synthetic loads and simplified environments, mainly evaluating platform scheduling and collaboration capabilities, and did not measure the actual execution capability of individual tasks.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

In the Router layer stress test, when the Trigger arrival rate increased from about 1 per minute to about 40, the average batch size increased from 1.0 to 2.8, the time taken per request increased from 2.5 seconds to 3.9 seconds, but the inference time allocated per decision decreased from 6.4 seconds to 1.6 seconds. Based on this, the team believes that during event bursts, the Router can improve effective throughput by merging backlogged events.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

The article also pointed out that this architecture still has challenges to be addressed, including how to scale the Router layer with task size, long-term memory spanning months or even longer timeframes, human-computer collaboration methods outside of Sessions, A2A communication between different Agent instances, and security and privacy protection. Especially when cloud Agents access local devices, issues such as prompt injection, local credentials leakage, and cross-device permission risks need further handling.

DeepSeek proposes Super-agent architecture with Router-Worker dual layers for more autonomous agents

The DeepSeek team stated that internal Super-agent instances built based on the aforementioned concepts and DeepSeek Harness have been continuously running, participating in issue reproduction and fixing, code review responses, and the self-iteration of DeepSeek Harness.

Additionally, they presented a prototype implementation named dsh-superagent-prototype in the article, but explicitly noted that this plugin is used to validate the technical route and is not an official DeepSeek model.

References:

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry