大家好 我是为什么叫QQ

Hello everyone, I'm WhyQQ.

DeepSeek刚公开了一个Super-agent研究原型

DeepSeek recently unveiled an experimental Super-Agent research prototype.

我最在意的 倒是它想把Session

What intrigued me most is its deliberate effort to make the session window

从你的产品界面里藏起来

completely vanish from your product interface.

你不用再记住 哪个窗口修Bug 哪个窗口跑评审

You no longer need to remember which tab is fixing a bug, which is reviewing PRs,

哪个窗口还在等集成

or which is waiting on CI integration.

消息从IM和GitHub进来

Incoming messages flow from Slack, Lark, and GitHub;

系统自己找人手 续上下文 交付结果

the system assigns workers, resumes context, and ships results autonomously.

文章给出一个观察

The research paper shares an insightful observation:

人通常只能同时跟进三到五个活跃Session

a human engineer can comfortably track only three to five active session tabs simultaneously.

模型继续变强 人却可能卡在窗口切换上

As underlying models grow more capable, human cognitive bandwidth gets choked by context-switching across windows.

对程序员来说 这个痛点很具体

For developers, this bottleneck is painfully real:

你开的Agent越多 越像在做人工调度

the more coding agents you spin up, the more your workday devolves into manual task dispatching.

你可能会问 多开几个Subagent

You might wonder: can't we just spin up a few subagents

再配一条优先委派的提示词 能不能解决？

and add an aggressive delegation prompt to solve this?

小规模任务可以

For simple toy tasks, that works fine.

消息一多 主Session仍要接每条消息

As traffic spikes, the primary session must still ingest every single message,

判断归谁 再转发后续

decide routing ownership, and forward subsequent updates.

它一边跟人沟通 一边维护任务状态

It juggles human chit-chat while tracking low-level task state,

还要防止新输入打断当前工作

all while guarding running jobs against untimely interruptions.

上下文中混着闲聊 评审 CI和代码 窗口越来越挤

Prompt contexts become polluted with casual banter, PR reviews, CI traces, and diffs.

任务之间的隔离 全靠模型自觉遵守提示词

Isolation between distinct tasks hinges entirely on the LLM's willingness to follow system prompts.

DeepSeek的实验里 仅改一句委派提示

In DeepSeek's ablation tests, tweaking a single phrasing in the delegation prompt

完成时间和长尾就明显变化

caused completion latency and long-tail variance to swing wildly.

问题已经落到系统结构上 单靠提示词很难稳定

When a failure mode is structural, system prompts alone cannot guarantee reliability.

先把三个对象分清

First, let's distinguish three separate components:

设计文章描述团队内部的

the research paper describes the internal architecture

Super- agent实例和后续方向

of DeepSeek's internal Super-Agent instance and future roadmap;

GitHub仓库是公开研究原型 只实现其中一部分

the GitHub repo is a scoped research prototype implementing a subset of those concepts;

DeepSeek Harness则是底层运行环境

and DeepSeek Harness provides the underlying runtime scaffolding.

公开版面向一个GitHub仓库

The open-source prototype attaches to a single GitHub repository.

可接入飞书 允许的人发起请求后

When an authorized team member issues an instruction via Lark or IM,

系统建立工作项 独立Session和目录

the system provisions a work item, an isolated session, and a sandboxed workspace directory,

再回到原会话交付PR

eventually returning to the original chat thread to hand over the finished pull request.

README明确写着 它仅供研究参考

The README explicitly notes that it is published strictly for academic research reference.

公开代码与内部版本 文章实现都有差异

The open-source code differs from internal production deployments and paper benchmarks;

后面谈到的Watchdog和云端解耦

concepts like external Watchdogs and decoupled cloud loops

按架构设计理解 别当作安装后就有的成品

should be studied as architectural blueprints rather than turnkey out-of-the-box features.

如果你也在多个会话间来回切 把数量打在公屏上

If you find yourself constantly tab-switching across agent sessions, let me know in the comments.

第一层是常驻Router

The top layer is a persistent, long-running Router.

它统一接IM GitHub 邮件等入口

It serves as the single unified ingress across IM channels, GitHub events, and emails,

维护跨任务上下文 决定新事件交给谁

maintains cross-task state, and decides which worker receives new incoming events.

它更像团队里的值班协调人

Think of it as a team's on-call coordinator:

只处理高层关联 任务分派和对外回复

it handles high-level triage, task assignment, and external status broadcasts—

甚至可以不配Bash

and doesn't even need terminal shell execution privileges.

文章建议用较轻模型承担这层工作

The paper suggests deploying a lighter, faster model for this routing tier,

具体执行交给Worker

delegating deep execution down to specialized Workers.

各自拥有隔离上下文 工具和目录

Each Worker possesses its own isolated context window, toolset, and workspace directory.

Worker不直接面对人 关键结论回到Router

Workers never converse directly with end users; their milestone outputs return to the Router,

由Router统一回复

which synthesizes a coherent update to the user.

这样 用户面对一个连续对象

This presents a single continuous conversation partner to the human engineer,

底层却能并行推进多个任务

while the underlying engine advances multiple tasks in parallel.

Session留在实现层

Session lifecycle is relegated to an implementation detail,

不再占用你的项目管理注意力

freeing your cognitive bandwidth from tedious multi-window orchestration.

Worker和常见Agent Team看着相似 职责边界不同

While Workers may look like standard Agent Teams, their responsibility boundaries differ sharply.

Team Lead常要理解一个复杂任务的细节

A conventional Team Lead must grasp intricate sub-task nuances

再协调紧密耦合的步骤

and tightly coordinate sequential dependencies.

Router只管任务之间的关系 不介入单任务内部

The Router only manages cross-task relationships without meddling in single-task execution internals.

Worker拿到完整任务后

Once a Worker accepts a self-contained assignment,

可以继续调用自己的Subagent

it is free to spin up its own subagents

甚至组织一个Agent Team

or orchestrate an internal multi-agent squad.

这个嵌套关系很值得借鉴

This hierarchical nesting is a design pattern worth adopting:

你把独立工作项交给Worker

assign discrete work items to Workers,

把步骤协作留在Worker内部

confine step-by-step coordination within each Worker's sandbox,

跨项目共享的是事件和授权

and share only events and security mandates globally.

执行细节无需塞回全局窗口

Granular execution logs never pollute the global routing window,

Router因此能保持轻量

allowing the Router to remain lean and fast,

Worker也能选用更强的模型 互不拖慢

while Workers leverage flagship frontier models without bottlenecking one another.

主动接入群聊后 噪音会先到

Connecting an agent directly to team chat exposes it to overwhelming conversational noise.

文章用Fast Path和Slow Path处理这件事

The paper resolves this via a dual-stream Fast Path and Slow Path pipeline.

程序规则先查来源 群组 白名单和明确请求

The deterministic Fast Path checks event origins, channel whitelists, and explicit mention syntax,

直接挡掉无关事件

filtering out irrelevant chat noise at near-zero compute cost.

通过准入的候选 再交给LLM判断语义

Only candidate events passing admission reach the Slow Path LLM for semantic evaluation.

比如群里吐槽界面 可能只需记录；

For example, a casual UI complaint might simply be logged,

一条可复现的缺陷反馈 才可能触发后台复现

while an actionable bug report with reproduction steps triggers background triage.

确定性规则便宜 稳定 也方便审计

Deterministic rules are cheap, stable, and straightforward to audit,

模型只处理需要理解的少数候选

reserving expensive inference tokens for genuine comprehension tasks.

你做类似系统时

When architecting an autonomous agent ingress,

可以先统计每天事件量和真正接单量

measure your daily inbound event volume against actual actionable tasks

再决定哪些判断值得固化为代码

before deciding which heuristics should be hardcoded in software.

把你的过滤策略打在公屏上吧

What filtering strategies do you use? Let me know in the comments.

新任务仍由Router冷启动

New tasks are cold-started through the central Router.

Worker创建后 会订阅对应的Thread

Once initialized, the Worker subscribes directly to the relevant communication Thread.

同一个PR的CI结果 评审评论

Subsequent CI build results, review comments,

或同一个IM话题的回复 后续直接投给这个Worker

or follow-up thread messages route directly to that Worker

不再要求Router重新分派

without requiring the Router to re-evaluate dispatch decisions.

订阅可以是多对多

Subscription topologies can be many-to-many:

相关Worker能收到同一线程的新事件

collaborating Workers receive live updates from the same thread,

输出仍回到Router 保持全局记忆和统一表达

while final outputs funnel through the Router to preserve global memory and coherent voice.

这里的收益主要是省路由决策和Token

The primary benefit here is slashing redundant routing hops and prompt tokens;

完成速度未必提升

end-to-end completion latency doesn't necessarily shrink.

这个区分很关键 直通优化的是消息路径

This distinction is crucial: subscription bypass streamlines message routing,

Worker执行任务的时间并没有因此被压缩

but the physical execution time spent inside the Worker remains unchanged.

你会优先省延迟 还是省Token？

Would you prioritize cutting latency or saving tokens? Share your take.

事件进了Session 还要控制上下文密度

Once an event enters a session, managing context density is paramount.

Inline& Spill把高概率用到的内容直接放入窗口

The Inline & Spill mechanism keeps high-probability data inlined in the active prompt window,

长文只留摘要 开头和读取入口

while spilling verbose text into persistent storage with short summaries and pointer URIs.

每个Trigger Thread Worker都有ID

Every Trigger, Thread, and Worker carries a unique UUID

并带上关联对象ID

linked to its parent entities.

模型需要前情时 从当前消息读到Thread

When an agent requires historical context, it navigates from the current event up the Thread hierarchy,

再沿引用找到Worker PR或更早事件

resolving references back to specific Worker runs, PRs, or earlier triggers.

这就是Chain of Recall 它像按需展开的引用图

This is 'Chain of Recall'—an on-demand citation graph

工程上比反复贴全量历史更可控

that is vastly more manageable than repeatedly dumping monolithic raw logs,

也给缓存和权限检查留下明确边界

while establishing unambiguous boundaries for caching and authorization audits.

你可以从统一ID和读取工具开始做

You can start incrementally by assigning uniform UUIDs and reading utilities,

无需一次建成完整记忆系统

without needing to architect a monolithic memory engine all at once.

经典ReAct Loop调用同步run()

The classic ReAct agent loop calls synchronous tool operations via blocking run() invocations.

长命令没结束 会话就难处理新输入

If a long compilation or test suite is executing, the session cannot handle incoming user input.

文章把耗时工具改成start()和stop()

DeepSeek refactors long-running tools into decoupled start() and stop() primitives.

start()立即返回Started

start() immediately acknowledges with a Started signal,

后台继续执行 完成结果作为新事件投回会话

running in the background and returning the final result as a standalone event to the session.

用户此时能补充约束 叫停任务

In the meantime, the user can supply new constraints or cancel the task,

或发起另一项工作

or initiate a completely separate work item.

公开原型里 前台Shell运行五秒后也会转入后台

In the open prototype, foreground shell tasks switch to the background after 5 seconds,

并给人工后续消息保留额外槽位

reserving dedicated slots for subsequent human follow-ups.

文章同时注明 异步任务机制已在Harness实现

The paper notes that this asynchronous task engine is already implemented in DeepSeek Harness,

尚未默认开启

though not yet enabled by default.

你看两条时间线就能发现 差别在后台运行期间

Comparing the two timelines reveals that the fundamental difference lies in background execution:

前台还能不能继续接收消息

whether the foreground interface remains responsive to new incoming messages.

多信道交互还需要控制说话方式

Multi-channel agent deployments also demand adaptive communication protocols.

IM里 没人想看Worker连续刷几十次工具调用

In team chat channels, no one wants to see an agent spam dozens of raw tool invocation dumps.

Router只给简短结论 待决策问题和验收通知

The Router broadcasts only concise summaries, pending human-in-the-loop decisions, and delivery notices.

文章验证过单次回复不超过100词的规则

The paper validates a strict heuristic capping individual chat replies at under 100 words.

到了GitHub或邮件 交付可以更完整

On GitHub PRs or email threads, however, deliverables can be comprehensive,

写清问题背景 改动范围和测试结果

detailing problem context, diff blast radiuses, and verification test logs.

执行过程留在Worker 渠道规范放在Router

Keep noisy execution traces inside Workers, and enforce channel-specific communication formats at the Router.

这个分工能减少打扰

This division prevents cognitive fatigue

也避免同一段长回复被机械复制到所有平台

and stops verbose logs from being mechanically pasted across platforms.

你评估Agent时 可以记录每完成一项任务

When evaluating AI agents, track how many times it interrupts a human

它打断了人多少次

to complete a single production task.

这项指标比在线时长更接近真实体验

Interrupt frequency reflects real-world developer experience far better than uptime vanity metrics.

持久在线涉及三个分开的部分

True 24/7 autonomous availability requires decoupling three distinct tiers:

云端Agent Loop负责计算

a cloud-hosted Agent Loop running inference,

Session数据库保存状态

a persistent session database safeguarding state,

本地或云端daemon负责执行命令

and a local or remote host daemon executing physical shell commands.

开发机不用把未提交修改

Developer workstations don't need to mirror uncommitted local changes,

私有凭据和内网工具链完整复制到云端

private credentials, or intranet toolchains up to third-party cloud environments.

设备暂时离线时 会话状态仍在；

If your laptop sleeps, session state remains intact in the cloud DB;

重新上线后 任务可以继续 这就是BYOM的吸引力

once reconnected, jobs resume seamlessly—that is the beauty of Bring Your Own Machine (BYOM).

不过 云端模型一旦能驱动你的机器

However, granting cloud LLMs execution capabilities on your physical machine

提示词注入和凭据泄露的影响也会扩大

drastically amplifies the blast radius of prompt injections and credential leaks.

架构图上的轻量daemon

The lightweight host daemon shown in architecture diagrams

落地时需要认真做认证 权限和审计

demands rigorous mutual authentication, scoped RBAC, and tamper-proof audit trails.

别把它能连上 当作可以全权操作

Never conflate network connectivity with carte-blanche execution privileges.

长期运行还会遇到另一类故障

Long-running autonomous agents also encounter subtle operational failures:

进程活着 Worker却可能反复读同一个文件

a process remains healthy while the Worker gets stuck reading the same file,

连续修同一个失败测试 或逐渐偏离目标

repeatedly tweaking the same failing test in circles, or drifting from the original objective.

外部Watchdog Session由定时消息驱动

An external Watchdog Session driven by periodic cron triggers

检查接口报错

monitors for recurring API error bursts,

无效循环 缺少实质进展和目标偏移

infinite execution loops, stagnant progress, and semantic drift,

再尝试纠正或恢复

before attempting automated remediation or session reset.

它位于主系统之外

Operating strictly outside the primary agent loop,

避免主Loop自己卡住后连自检也一起停掉

it prevents a frozen main loop from silencing internal health checks.

这个思路可以先用普通监控实现

You can implement this today using standard observability tooling:

记录每个Worker最近的状态变化

track timestamped state transitions,

工具调用和交付物

tool call frequencies, and concrete artifact outputs for every active Worker.

没有进展信号 就别只用心跳判断健康

Without forward progress signals, a bare TCP heartbeat is never proof of health.

主动性必须配授权

Autonomous proactivity must be tethered to strict authorization guardrails.

公开原型有off observe auto三种接受模式

The research prototype implements three execution modes: off, observe, and auto.

默认observe

In the default 'observe' mode,

匹配规则的推断任务只记录would accept

tasks triggered by heuristics are merely logged as 'would accept'

不真正执行

without mutating live production code.

人的明确请求和机器人的候选事件也被分开

Explicit human requests are segregated from agent-inferred background events.

Mandate会记录授权范围

Security Mandates capture permissible file scopes,

命中规则 策略版本和来源身份

matched rule triggers, policy versions, and verified caller identities.

还有一条安全提醒很直接

One blunt security reality must be stated:

Worker可用运行账户的权限执行命令

Workers execute shell commands with the full privileges of their host OS account.

目录封装和禁用命令只是默认约束

Directory chroots and command blocklists are merely soft baseline constraints.

试用时请用独立账户

Always test agents under dedicated service accounts

最小权限 隔离环境和分支保护

with least-privilege policies, containerized sandboxes, and branch protection rules.

安全需要系统强制 提示词只负责行为引导

Security must be strictly enforced by host systems—prompts can only guide behavioral etiquette.

文章的消融实验值得看 也要看懂口径

The paper's ablation benchmarks are compelling, but their methodology must be understood.

每轮约300条群聊和GitHub CI消息

Each test round simulated roughly 300 group chat and GitHub CI events,

其中约20条是工作请求

where only about 20 represented actionable work requests

其余包括闲聊 跟进和叫停

and the remainder consisted of chatter, check-ins, and abort signals.

编码被一个只消耗时间的工具替代

Code generation was mocked by a deterministic time-delay tool,

消息间隔和工作时长按四分之一缩放

message intervals and task durations were scaled down by 4x,

模型调用延迟保留真实值 每种方案跑5次

while LLM API latency was preserved over 5 benchmark runs per configuration.

所以它测的是调度与协作 不测代码正确率

It measures orchestration overhead and multi-agent coordination—not synthesized code correctness

也没有比较真实项目中的修复质量

or real-world patch survival rates.

实验里Super- agent的P50和P95

Under this benchmark, the Super-Agent's P50 and P95

完成时间分别为12.4和15. 7分钟

completion latencies clocked in at 12.4 and 15.7 minutes respectively.

先记住这两个数

Keep those baseline figures in mind.

去掉订阅直通后 P50和P95是12.6和15.5分钟

Disabling thread subscription bypass yielded P50 and P95 latencies of 12.6 and 15.5 minutes—

差别很小；

virtually unchanged on elapsed time.

cache-read tokens却从2323万涨到3177万

However, prompt cache-read tokens surged from 23.23M to 31.77M—

约增37% Router决策约增30%

a ~37% increase in token consumption and a ~30% spike in Router decisions.

Router还有批处理兜底

The Router also features batch deduplication fallback:

事件速率升高时 它会合并积压的Trigger

during sudden traffic bursts, it coalesces backlogged triggers into grouped inferences,

所以负载增加后 也没有撞上单Router的吞吐上限

preventing a single Router node from choking on throughput limits.

单Session加Subagent在中性

By comparison, a single-session architecture with subagents under neutral prompts

提示下是20. 4和29. 5分钟

registered P50 and P95 latencies of 20.4 and 29.5 minutes;

优先委派提示下是13.8和29.0分钟

under aggressive delegation prompts, it scored 13.8 and 29.0 minutes.

后者的长尾来自某次未委派

The severe long-tail lag in the latter stemmed from occasional prompt delegation misses.

Token口径也要看清

Token economics must also be weighed accurately:

Super-agent总读缓存高于单Session

the Super-Agent consumes more total cache reads than single-session models,

它用更多上下文换更稳的并行调度

trading broader context memory for bulletproof parallel scheduling,

直通再帮忙省掉重复路由

while thread bypass claws back redundant routing overhead.

这里没有全维度获胜 只有具体取舍

There is no free lunch in AI architecture—only deliberate engineering trade-offs.

很多人会觉得

Many engineers assume

多Agent系统就该继续加角色 加沟通 加协调

scaling multi-agent systems means endlessly piling on more roles, protocols, and chatty debates.

看到这里 我的建议相反

Having analyzed this design, my advice is the opposite:

先把人的交互收拢到一个入口 把独立任务隔离

first consolidate human interaction into a single ingress, isolate independent tasks into sandboxes,

把已知路由写成程序 再让模型处理语义和异常

encode deterministic routing in code, and only then deploy LLMs for semantic nuance and anomaly handling.

你可以用四个问题检查自己的系统：

Audit your own multi-agent architecture with these four questions:

新消息会不会阻塞旧任务

1. Do incoming events block running background jobs?

后续事件要重复路由几次 长文是否反复进入窗口

2. How many times are thread follow-ups re-routed, and does verbose text pollute prompts?

授权能否脱离提示词被强制执行

3. Can security boundaries be strictly enforced independent of LLM compliance?

四项都有答案后 再考虑更复杂的协作拓扑

Once you have solid answers to these, only then explore complex multi-agent topologies.

架构收益来自减少无用工作

Architectural ROI comes from eliminating wasted work—

角色数量本身提供不了这种保证

simply inflating agent headcounts provides zero guarantees.

接下来值得观察的

The most compelling frontiers to watch next

是授权审计 跨设备安全和长程记忆

are fine-grained authorization audits, cross-device execution safety, and durable long-term memory.

公开原型也还没覆盖全部设计

The open-source prototype represents only a foundational sliver of the full architecture.

你如果想动手 建议从非核心仓库开始

If you plan to experiment, start on a non-critical repository:

只接明确请求 先跑observe

listen only to explicit mentions, run in 'observe' mode first,

记录每次接单 打断和交付

log every accepted task, human interruption, and deliverable,

再逐步开放一条窄规则

and gradually open up narrow automated rules one by one.

这个方向真正节省的是人的调度成本

The true prize of this paradigm is reclaiming human attention from context orchestration,

让工程师回到方案选择和结果验收

freeing engineers to focus on high-leverage architectural choices and outcome verification.

你更愿意把哪类任务交给这样的Agent

Which workflows would you trust to an autonomous agent like this—

缺陷复现 CI跟进 还是PR评审？欢迎评论区聊聊

bug reproduction, CI follow-up, or PR reviews? Let's discuss in the comments.

感谢观看 别忘了一键三连

Thanks for watching! Don't forget to like, subscribe, and share.

会话该由系统管理

Sessions belong to system infrastructure;

你的注意力值得留给更难的问题

your attention belongs on the hardest engineering problems.

我是为什么叫QQ 我们下期再见

I'm WhyQQ—see you in the next episode.