DeepSeek 重磅发布全新 Agent 架构：AI 开始指挥 AI 干活
当程序员还在群聊里讨论 Bug 怎么修时，AI 可能已经默默完成了问题复现，甚至把修复代码提交到了 GitHub。
就在刚刚，DeepSeek Harness 团队发表技术文章《Super-agent 设计探索：迈向更自主的 Agent 系统》，详细介绍了他们正在探索的 Super-agent 架构。
按照 DeepSeek 的设想，未来的 AI Agent 将不再局限于单个聊天窗口，而是能够持续在线、同时处理多项任务，并根据外部环境变化主动采取行动。
更值得关注的是，DeepSeek 已经基于 DeepSeek Harness 构建了一套内部 Super-agent 实例，并将其用于日常研发协作。
例如，当团队成员在群聊中反馈软件问题时，即使没有人主动 @ AI，它也能识别潜在缺陷，在后台尝试复现问题、寻找修复方案，甚至直接提交 GitHub Pull Request，等待开发人员审查。
同一项开发任务还能在群聊和 GitHub 之间连续推进。开发者先在群聊中确定方案，随后又在 GitHub 评论中修改要求，Agent 能够理解前后变化，并调整执行方向。
不过，Super-agent 目前仍是一套面向工程实践的架构探索，部分设计已通过开源原型实现，尚非 DeepSeek 正式推出的独立产品。
AI Agent 的下一个瓶颈，可能是人类自己
过去一年，以 Coding Agent 为代表的 AI 工具已经能够独立阅读代码、定位 Bug、运行测试，甚至完成复杂的软件开发任务。
然而，随着模型能力越来越强，另一个问题也开始显现。
AI 能够同时完成的任务越来越多，人类管理这些任务的注意力却没有同步增长。
目前，大多数 Agent 仍以 Session，也就是会话为基本工作单位。用户需要在不同窗口之间反复切换，记住各个任务的上下文，并持续跟踪执行进度。
DeepSeek 团队提到，根据他们的观察，一个人通常只能同时跟进 3 到 5 个活跃 Session。
即使通过 CLI、MCP 等工具让 Agent 接入更多平台，仍然难以解决单会话串行执行、多任务上下文混乱，以及本地设备离线后任务中断等问题。
为此，DeepSeek 提出了 Super-agent，希望让 Agent 从单次会话中解放出来，成为一个能够长期运行、跨平台保持上下文、并行推进任务的连续实体。
按照文章中的定义，Super-agent 需要具备四项能力，包括跨会话的连续交互、多任务高效并行、主动响应环境事件，以及长期在线和异常恢复。
换句话说，人类可以更多地负责确定目标和验收结果，具体的任务调度与执行过程则交给 AI 管理。
一个 AI 当主管，多个 AI 同时干活
为了实现这一目标，DeepSeek 设计了一套由 Router 和 Worker 组成的双层架构。
可以将其理解为一个由 AI 组成的虚拟团队，Router 负责统筹协调，Worker 负责具体执行。
Router 是持续运行的全局协调节点，统一接收来自群聊、GitHub、邮件等不同平台的信息，判断任务应该如何分配，并维护跨任务上下文。
真正执行任务的则是 Worker。每个 Worker 拥有独立的上下文和工具能力，可以分别负责问题复现、代码审查、功能开发等任务，必要时还能继续调用 Subagent。
任务完成后，Worker 将关键结果汇报给 Router，再由后者统一与用户沟通。
对于用户而言，面对的始终是同一个 AI 助手；对于系统而言，背后可能已经有多个独立 Agent 同时工作。
为了减少协调开销，DeepSeek 还引入订阅机制。当某个 Worker 接手 GitHub Pull Request 后，后续评论和状态变化可以直接发送给它，无须每次经过 Router 重新分配。
Router 主要负责信息理解和任务分派，因此相对轻量的模型也可能胜任，能力更强的模型则用于复杂任务执行。
DeepSeek 认为，传统 Agent 让用户手动管理 Session，很大程度上延续了底层大模型 API 的设计习惯。随着 Agent 能力提升，Session 应该逐渐成为系统内部的技术细节。
没人 @ 它，AI 也能主动修 Bug
相比多任务并行，Super-agent 更值得关注的能力，是主动寻找值得处理的事情。
传统 AI 助手通常需要用户先提出问题，再调用工具完成任务。DeepSeek 希望改变这种工作方式，让 Agent 直接根据外部环境中的事件采取行动。
群聊消息、邮件、日程提醒、GitHub 评论以及 CI 测试结果，都可以成为触发 Agent 行动的事件，技术上称为 Trigger。
考虑到绝大多数群聊消息并不需要 AI 介入，DeepSeek 设计了两级过滤机制，由程序规则快速排除无关消息，再交给大模型结合上下文判断是否需要行动。
更进一步，DeepSeek 还引入了计算机系统中的投机执行思想。
所谓投机执行，就是在尚未确定结果是否会被采用时，提前利用后台算力进行尝试，有价值的结果再提交给用户，没有价值则直接放弃。
例如，DeepSeek 在内部用户群中配置了一项「静默复现并修复」策略。
当有人反馈 Bug 或讨论修改方案时，Super-agent 可以提前在后台复现问题、分析原因，并寻找修复方案。
只有确认问题存在，并找到相对简单的修复方法后，AI 才会把结果提交给开发人员，否则全程保持安静。
DeepSeek 展示的案例中，团队成员还在群聊里讨论解决方案，Super-agent 就已经在后台提交了修复 Pull Request。
类似机制也可以应用于个人助手，例如收到会议邀请后，AI 可以提前核对日程、搜集背景资料，并准备邮件回复草稿，等待用户确认。
电脑关机，AI 也能继续工作
除了主动性，DeepSeek 还尝试解决传统 Agent 的响应阻塞和设备依赖问题。
在常见的 ReAct Loop 中，模型发起耗时工具调用后，需要等待结果返回，期间可能无法及时响应新的指令。
DeepSeek 因此设计了异步 ReAct Loop，让耗时任务转入后台执行，前台会话继续接收用户消息，任务完成后再将结果发送回来。
AI 可以一边执行复杂任务，一边接受新指令，用户还能在执行过程中要求它改变方向或终止任务。
该机制已经在 DeepSeek Harness 中实现，但尚未默认开启。
与此同时，DeepSeek 借鉴存算分离的思路，将 Agent 的执行循环放在云端，将会话状态保存到数据库，而代码仓库、文件和工具链等工作环境仍然可以留在用户设备上。
本地只需部署一个轻量级 daemon，负责接收云端命令并返回结果。即使电脑暂时离线，相关任务也可以在设备重新上线后继续执行。
此外，DeepSeek 还引入独立的 Watchdog 机制，相当于给 Agent 安排了一个 AI 监督员，负责检查任务是否取得进展、是否陷入无效循环，并在发现异常时尝试修复。
实测效率提升，但也更费 Token
为了验证新架构，DeepSeek 团队进行了一组初步实验。
实验模拟群聊和 GitHub CI 消息流，单次包含约 300 条消息，其中约 20 条是真正的工作请求。研究人员将 Super-agent 与传统单 Session 加 Subagent 的方案进行对比，每种方案重复运行五次。
结果显示，Super-agent 的任务完成时间 p50 为 12.4 分钟，p95 为 15.7 分钟；采用中性提示词的传统单 Session 方案，p95 则达到 29.5 分钟。
差异主要来自任务调度。传统方案虽然支持 Subagent，但模型是否真正分派任务，仍取决于提示词和临场判断。
在中性提示配置下，五次实验中模型一次都没有调用 Subagent，所有任务都在主会话中执行，不断受到新消息干扰。
Super-agent 则通过双层架构，让 Router 专注任务分派，Worker 独立执行，减少了这种不确定性。
不过，效率提升也有代价。由于需要维护多个独立会话，Super-agent 消耗的缓存读取 Token 高于单 Session。
取消事件直通机制后，Router 需要做出的决定增加约 30%，总缓存读取量从 2323 万增加至 3177 万 Token，增幅约 37%。
另外，当事件到达速率从每分钟约 1 条升至 40 条时，Router 可以通过批量处理提高吞吐量，单次请求耗时仅从 2.5 秒增加至 3.9 秒。
需要注意的是，实验使用合成消息和模拟耗时工具，主要验证任务调度与协作能力，并未直接评估真实编码质量。
DeepSeek 想让 AI 成为真正的数字同事
尽管 DeepSeek 已经在内部运行 Super-agent，但距离大规模应用仍有不少挑战。
首先是长期记忆。当任务积累到数周甚至数月后，如何准确恢复历史上下文，仍然是一个难题。
其次是安全与隐私。由于 Super-agent 会持续接收外部消息，并通过云端访问本地设备，一旦遭遇提示词注入或服务端入侵，就可能威胁代码仓库、敏感数据和本地凭据。
此外，当 Agent 能够同时推进大量任务时，用户应该如何查看进度、介入执行，以及决定哪些操作必须人工批准，也需要重新设计。
目前，DeepSeek 已将部分设计实现为 dsh-superagent-prototype 原型插件，相关架构仍在持续探索。
值得注意的是，DeepSeek 团队在文章结尾坦言，他们并没有提出全新的理论，而是将任务调度、缓存管理、事件驱动、存算分离等经典计算机系统思想应用到了 Agent 架构中。
但随着模型完成单个任务越来越少需要人工干预，人类能够同时管理多少个 AI，正在成为新的问题。
过去，我们需要不断向 AI 发送指令，告诉它下一步做什么。按照 Super-agent 的设想，未来人类可以更多地负责确定目标、做出决策和验收成果，让 AI 自己发现问题、协调任务并完成执行。
或许未来真正成熟的 AI 助手，不会一直待在聊天窗口里等待提问，而是像一位能够长期参与工作的数字同事，在你没有主动找它的时候，也知道自己应该做些什么。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
DeepSeek weighs publish brand-new Agent structure: AI starts commanding AI working
When the programmer was still talking about Bug's repair, AI might have finished the problem silently and even sent the repair code to GitHub.
Just recently, the DeepSeek Harness team published a technical article, Super-agent Design Exploration: Towards a more autonomous Agent System, detailing the Super-agent architecture they are exploring.
As DeepSeek envisages, the future AI Agent will no longer be limited to a single chat window, but will be able to handle multiple tasks online on a continuous basis and on its own initiative in response to changes in the external environment.
More noteworthy is the fact that DeepSeek has built a set of internal Super-agent examples based on DeepSeek Harness and used them for daily research and development collaboration.
For example, when team members talk about feedback software issues in groups, even if no one takes the initiative @AI, it can identify potential deficiencies, try to repeat problems in the backstage, look for restoration options, and even submit them directly to GitHub Pull Request, pending review by developers.
The same development mission can also advance between the group chat and GitHub. The developers first define the program in the group chat and then modify the requirements in GitHub comments, so that Agent can understand the changes and reorient the implementation.
However, Super-agent is still a structured exploration for engineering practices, some of which have been implemented through the open source prototype and are not yet independent products officially launched by DeepSeek.
The next bottleneck for AI Agent is probably human.
Over the past year, the AI tool represented by Coding Agent has been able to read code, locate Bug, run tests and even complete complex software development tasks.
However, as modelling capacity grows, another problem begins to emerge.
AI has been able to accomplish more and more tasks at the same time, and human attention to managing these tasks has not increased simultaneously.
Currently, most Agents are still based on Ssession, which is a session. Users need to switch from window to window, remember context for tasks, and keep track of progress.
The DeepSeek team mentioned that according to their observations, a person can usually follow only three to five active sessions at the same time.
Even if Agent were to have access to more platforms through tools such as CLI, MCP, it would still be difficult to address single-chamber serial execution, multi-task context confusion and disruption of local equipment after offline.
To this end, DeepSeek proposed Super-agent, hoping that Agent would be liberated from a single session and become a continuous entity capable of operating on a long-term basis, maintaining context across the platform and advancing tasks in parallel.
As defined in the article, Super-agent requires four capabilities, including continuous cross-conference interaction, multi-task efficient parallels, proactive response to environmental events, and long-term online and abnormal recovery.
In other words, humans can be more responsible for setting targets and results of acceptance check, while specific tasking and execution processes are managed by AI.
One AI as supervisor, several AIs working at the same time
In order to achieve this goal, DeepSeek designed a two-tiered architecture consisting of Router and Walker.
It can be understood as a virtual team of AI, Router is in charge of coordination and Worker is in charge of implementation.
Router is an ongoing global coordination node that integrates information from different platforms, such as community chats, GitHub and mail, and determines how tasks should be allocated and maintains cross-mission context.
It's Walker who's really on the job. Each Worker has a stand-alone context and tool capability, which can be responsible for issues recurrence, code review, functional development and, if necessary, continued to call Subagent.
Upon completion of the mission, Worker reports the key results to Router, who then communicates them uniformly with users.
For users, it's always the same AI assistant; for the system, there may already be more than one independent Agent working at the same time.
To reduce coordination costs, DeepSeek also introduced subscription mechanisms. When a Worker takes over GitHub Pull Request, subsequent comments and changes of status can be sent directly to it without having to be redistributed through Router each time.
Router is primarily responsible for information understanding and tasking, so that relatively lightweight models may also be competent and more capable models can be used for complex task execution.
DeepSeek argues that traditional Agent allows users to manage the session manually, largely continuing the design habits of the bottom LLM API. As Agent's ability increases, the session should become a technical detail within the system.
@ It, AI can fix Bug
Super-agent's ability is more of a concern than multitasking, and it's a proactive search for something worth dealing with.
Traditional AI assistants usually need users to ask questions and then call the tools to do their job. DeepSeek wants to change this way of working so that Agent can act directly on events in the outside environment.
Group chats, e-mails, calendar alerts, GitHub comments and CI tests can all be events that trigger Agent, technically known as Trigger.
Considering that the vast majority of groups talk about information without the need for AI intervention, DeepSeek designed a two-tier filter mechanism, with procedural rules quickly removing irrelevant information, before LLM, in conjunction with context, decides whether action is required.
To go further, DeepSeek has also introduced speculation in computer systems.
Speculation enforcement is an attempt to use back-office calculations in advance, where it has not yet been determined whether or not the results will be used, and the valuable results are then submitted to the users, with no value, or simply abandoned.
DeepSeek, for example, has a silent recovery and repair strategy in its internal user base.
When feedback is received from Bug or changes are discussed, Super-agent can recreate problems, analyse reasons and look for rehabilitation programmes in advance.
AI will submit the results to the developers only if the problem is confirmed and relatively simple methods of repair are found, otherwise the whole process will remain quiet.
In the case of DeepSeek, where team members are still discussing solutions in a group chat, Super-agent has already submitted the restoration of Pull Request to the backstage.
Similar mechanisms could also be applied to personal assistants, such as AI, which, upon receipt of an invitation to the meeting, could check the schedule in advance, gather background information and prepare draft mail responses pending confirmation by users.
Computer shut-down, AI can keep working.
In addition to initiative, DeepSeek tried to solve the problem of traditional Agent's response jams and equipment dependence.
In a common React Loop, it takes time for the model to start tool call to wait for the result to return and may not be able to respond to new instructions in a timely manner.
DeepSeek has therefore designed a rect Loop to move time-consuming tasks backstage, backstage sessions continue to receive user messages, and the results are sent back when the mission is completed.
AI can perform complex tasks while accepting new instructions, and users can also ask it to change direction or terminate missions in the course of implementation.
The mechanism has been implemented in DeepSeek Harness, but has not yet been launched by default.
At the same time, DeepSeek, drawing on the idea of storage separation, keeps the execution cycle of Agent on the cloud and saves the conversation status to the database, while work environments such as code warehouses, files and tool chains can still be left on user devices.
Only one lightweight daemon will be deployed locally to receive cloud orders and return results. Even if the computer is offline for the time being, the task may continue after the equipment has been re-loaded.
In addition, DeepSeek has introduced an independent Watchdog mechanism, which is tantamount to having an AI supervisor assigned to Agent to check whether the mission is moving forward, falling into an ineffective cycle and attempting to repair it when anomalies are discovered.
Increased efficiency, but more expensive.
In order to validate the new architecture, the DeepSeek team conducted a set of preliminary experiments.
Experimental simulation group chats and GitHub CI newsstreams containing about 300 messages, of which about 20 are genuine job requests. The researchers compared Super-agent with the traditional Ssession plus Subagent programmes, each running five times over.
The results show that the time for completion of the Super-agent task was 12.4 minutes and 15.7 minutes, while the traditional Ssession programme using neutral prompt was 29.5 minutes.
The main causes of the variance are mission movements. While the traditional programme supports Subagent, the actual assignment of tasks to the model depends on prompt and on the on-site judgement.
Under the neutral hint configuration, the models were not called at any time in five experiments, and all the tasks were performed in the main session and were constantly interrupted by new messages.
Super-agent reduced this uncertainty by focusing Router on task assignments and working independently through a two-tiered structure.
However, efficiency gains also come at a cost. The Super-agent consumes more cache reading token than single session due to the need to maintain multiple independent sessions.
With the elimination of the Incident Direct Access mechanism, Router ' s decision was increased by about 30 per cent, from 23.23 million to 317.7 million Token, an increase of about 37 per cent.
In addition, when the event's rate of arrival rises from about 1 to 40 per minute, Router can increase the amount of throughput by batch processing, and single requests take only 2.5 seconds to 3.9 seconds.
It is important to note that the experiment, which uses the synthesizing message and simulation time-consuming tool, primarily validates mission movement and collaboration capabilities and does not directly assess the quality of real codes.
DeepSeek wants to make AI a real digital colleague.
Although DeepSeek is already operating internally, there are still many challenges to large-scale applications.
First, long-term memory. The precise restoration of history context remains a challenge as the task accumulates for weeks or even months.
The second is security and privacy. As Super-agent receives external information on an ongoing basis and accesses local equipment through clouds, it is possible to threaten code warehouses, sensitive data and local documents in the event of prompt injections or service-end intrusion.
In addition, when Agent is able to advance a large number of tasks simultaneously, how users should view progress, intervene in implementation, and decide which operations must be manually approved, also requires redesign.
At present, DeepSeek has partially designed to be a prototype dsh-suberagent-prototype plugin, and the structure is still being explored.
It is worth noting that the DeepSeek team stated at the end of the article that they did not propose a completely new theory, but instead applied the classic computer system ideas of task movement, cache management, event drive, memory separation to the Agent structure.
But as models become less and less manual to perform individual tasks, the number of AIs that humans can manage at the same time is becoming a new issue.
In the past, we need to keep sending orders to AI to tell it what next step does. According to Super-agent ' s vision, future human beings can be more responsible for setting goals, making decisions and making acceptance check results, allowing AI to identify problems, coordinate tasks and implement them.
Perhaps the real mature AI assistant in the future will not stay in the chat window waiting to ask questions, but like a digital colleague who can work for a long time, knowing what to do when you don't look for it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.