背景
开发者日常工作涉及多个平台切换：在 IM 群聊讨论方案，到 GitHub 提交代码，再根据 PR 评论修改实现。传统 Coding Agent 将这些操作分散在独立会话中，用户需要手动维护上下文，在不同会话间转述任务进展。DeepSeek 团队观察到，一个人通常只能同时跟进 3 到 5 个活跃 Session，这限制了多任务协作效率。模型单项任务能力变强，人类注意力反而成为瓶颈。
Super-agent 解决的正是这一问题。用户面对同一个 Agent，从群聊决策到 GitHub PR 评审，系统自行管理底层多个会话。DeepSeek 团队内部实例中，Agent 能理解群聊里采用方案 A 和 GitHub 评论里改用方案 B 属于同一项工作，并据此调整执行方向。Agent 还引入投机执行：群里讨论 Bug 时，后台已着手复现和定位，找到可行修复方案后交给用户，未找到有效结果则保持安静。
核心原理
Super-agent 采用双层架构。上层 Router 统筹全局，负责接收消息、分配任务和维护上下文。下层 Worker 是一组 Agent 会话，负责具体执行，多个任务可同时推进。关键设计是订阅与直通机制：Worker 接手 GitHub PR 后直接订阅后续评论，新的修改意见无需经 Router 转发，直接交给对应 Worker 处理。老任务持续跟进，新任务及时分配。
消息过滤采用两级机制：程序规则先筛除无关消息，大模型再判断剩余事件是否需要行动，减少 Token 消耗。异步 ReAct Loop 让 AI 执行耗时任务时仍能响应新指令，无需等待上一任务完成。该机制已在 DeepSeek Harness 中实现，但默认未开启。
24 小时在线依赖计算、状态与执行环境分离。Agent 计算逻辑在云端运行，会话状态独立保存，本地设备保留工作环境。电脑离线时，云端 Agent 维持任务状态，设备重新上线后继续执行。Watchdog 机制作为独立监督 AI，既检测程序故障，也判断任务是否陷入无效循环或偏离目标，并尝试介入修复。
局限
效率提升伴随一定 Token 开销。取消事件直通机制后，Router 决策次数增加约 30%，读取缓存 Token 增加约 37%。合理的事件路由能降低运行成本，但多层架构本身会引入额外消耗。
初步实验在约 300 条消息、20 项工作请求的模拟环境中进行，每种配置重复测试 5 次。Super-agent 任务完成时间中位数 12.4 分钟，95 分位 15.7 分钟；传统单 Session 方案在中性提示下，95 分位达到 29.5 分钟。实验规模有限，更复杂场景下的表现有待验证。异步 ReAct Loop 和投机执行机制尚未默认开启，功能成熟度需进一步观察。
对开发者的意义
Super-agent 让 Agent 从单次会话限制中解脱，实现跨平台持续协作。开发者无需在多个会话间反复交代上下文，Agent 能跟踪从群聊讨论到代码提交的完整流程。主动响应和投机执行减少等待时间，Agent 在后台尝试解决方案，找到结果后直接呈现，未找到则不打扰。
并行执行让多个复杂任务同时推进，避免单个耗时任务阻塞后续工作。24 小时在线和自动恢复机制支持长期运行，遇到异常时尝试自行修复。这些特性正在将 Agent 从「等待人类逐项分配任务」推向「主动发现问题、持续推进工作」。
DeepSeek 公开架构思路，也为开发者探索类似系统提供了参考。GitHub 仓库已开放原型代码，开发者可以基于此进一步实验和改进。
原文地址：tools.wflynn.cn/aitools/tre…
关注公众号「YourTools」，每天推送 GitHub 趋势与开源工具速报：
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。
Background
Developer's day-to-day work involves multiple platform switches: discussion options in the IM group, submission codes in GitHub, then changes made in accordance with PR comments. The traditional Coding Act spreads these operations into separate sessions, and users need to maintain context manually and relay mission progress between sessions. The DeepSeek team observed that one person usually follows only three to five active sessions simultaneously, which limits the effectiveness of multi-task collaboration. Human attention has become a bottleneck when the individual tasks of the model have become more capable.
Super-agent solves the problem. Users face the same Agent, from group conversations to GitHub PR reviews, and the system manages its own bottom-up sessions. In the example of the DeepSeek team inside, Agent understands that the change of option B to option A in the group conversation and GitHub comments is the same job and that the implementation direction is aligned accordingly. Agent also introduced speculative execution: when Bug was discussed in the group, the backstage had begun to re-emerge and locate, to find a viable repair solution for the user, and to remain quiet if no effective results were found.
Core principles
Super-agent uses a double layer structure. Upper Router General, which is responsible for receiving information, assigning tasks and maintaining context. Worker is a set of Agent sessions that are actually being performed and multiple tasks can be advanced simultaneously. Key design is subscription and direct access: Worker takes over GitHub PR and subscribes directly to subsequent comments, and new changes need not be forwarded through Router and sent directly to the corresponding Worker. Old tasks follow up and new tasks are allocated in a timely manner.
The filtering of messages involves a two-tier mechanism: the rules of procedure remove irrelevant information and LLM then determine whether the remaining incidents require action to reduce token consumption. rect Loop allows AI to respond to new instructions while performing time-consuming tasks without waiting for the previous task. The mechanism was implemented in DeepSeek Harness, but was not opened by default.
24 Hours of online reliance on computing, state and execution environments. Agent calculates the logic of the cloud, saves the session status independently, and the local device retains the working environment. When the computer offline was in operation, the cloud Agent maintained its mission and the equipment was re-loaded and continued. The Watchdog mechanism, as an independent monitoring AI, both monitors process failures and determines whether the mission is in an ineffective cycle or de-targeting and attempts to intervene in repair.
Limits
Efficiency gains are accompanied by a certain Token cost. With the elimination of the incident straight-through mechanism, Router increased the number of decisions by about 30 per cent and cache Token by about 37 per cent. Reasonable event routing reduces operating costs, but multi-layer structures themselves introduce additional consumption.
The initial experiments were conducted in a simulation environment of approximately 300 messages, 20 job requests, and were repeated five times for each configuration. Medium number of Super-agent tasks completed 12.4 minutes, 95 minutes 15.7 minutes; traditional single-session programmes reach 29.5 minutes at a neutral tip. The scale of the experiment is limited and the performance in a more complex scenario needs to be validated. The rect Loop and speculative enforcement mechanisms have not yet been defaulted on, and functional maturity needs to be further observed.
Meaning to developers
Super-agent lets Agent out of a single session limit and achieves continuous cross-platform collaboration. Developers do not have to repeat context in multiple sessions, and Agent can track the entire process from group chat to code submission. Active response and speculative execution reduced the waiting time, and Agent tried solutions in the backstage, presenting them directly when results were found, without disturbing them.
Parallel implementation allows multiple complex tasks to advance at the same time and avoid individual time-consuming tasks blocking follow-up. The 24-hour online and automatic recovery mechanism supports long-term operation and attempts to repair itself in case of anomalies. These features are moving Agent from "waiting for the assignment of human tasks on a case-by-case basis" to "proactive identification of problems and continuous work."
DeepSeek's open architecture thinking also provides a reference for developers to explore similar systems. The GitHub repository has opened the prototype code, on which the developers can build further experiments and improvements.
Original address:tools.wflynn.cn/aitools/tre...
Following the WeChat Official Account "YourTools", GitHub's trend and open source tool is sent daily:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.