前言
最近B站科技区有个挺有意思的现象。
DeepSeek Harness（下面简称DSH）官方入驻B站，发了一条视频，24小时播放破了211万，账号一天涨了40万粉。
但真正让我停下来看了很久的，是评论区。
不是那种“沙发”“前排”的灌水。
有人在晒自己给DSH写的插件，有人在讨论某个插件的实现细节，有人在问“我这边十五个插件了，有没有比我还多的”。
还有人把自己女儿画的画做成了一个平台跳跃小游戏，拿了六千多个赞。
这不是官方组织的活动，这是一群人在自发地“造东西”。
我做了这么多年Java，见过开源项目火，见过框架出圈，但一个Agent框架的评论区变成开发者作品展，这件事本身比任何跑分都值得琢磨。
它说明一个问题：Agent的竞争，正在从“谁的模型更聪明”，转向“谁的生态更能让人愿意动手”。
一、先看看评论区里到底在晒什么？
DSH官方那条入驻视频，讲的其实是它的“创造模式”——在对话里说一句想要什么，它就把想法写成插件，写好直接能用。
视频里演示了倒计时、便签墙、沙盒游戏、Transformer的3D可视化。
但评论区晒出来的东西，比官方演示野得多。
有人做了dsh猫猫插件。
有人做了可视化的近现代史纲和毛中特，底下一排人求链接，说对考研复习有用。
有人做了手机远程连接工具，专门优化通知和UI适配。
还有眼尖的观众发现左侧会话列表里躺着一句“构建 AGI，务必做好”，截图出来刷了好多遍。
这些插件有一个共同点：它们不是“通用能力”，是“我的需求”。
专业开发者写的是基础设施——记忆、调度、远程连接。普通爱好者写的是“我每天用的时候觉得不方便的那一下”。
两边都在写，两边都在往同一个生态里扔东西。
B站上甚至已经有了完整的教程生态。
有人出了《DeepSeek Harness从入门到实战》的长视频，从底层原理讲到插件开发。
有人做了桌面端实测，手把手教你怎么把PDF阅读器做成DSH插件。
还有知乎专栏的实战教程，直接演示“我让DeepSeek自己写了一个arXiv搜索插件”。
一个框架能不能活，不看官方发多少文档，看有没有人愿意自发地写“第二篇教程”。
二、DeepSeek的Super-Agent在赌什么？
评论区热闹归热闹，但真正决定Agent能走多远的，是另一个问题：当一个人同时管五个Agent的时候，谁来管这五个Agent？
DeepSeek在10月31日发布的Super-Agent技术文章里，把这个痛点说得很直白。
当前Coding Agent大多以Session为中心，每个Session只管自己那一摊。
一个人通常只能同时跟进3到5个活跃Session。
AI越来越能干，人反而变成了“五个AI的项目经理”。
Super-Agent的解法是Router–Worker双层架构。
Router负责接收消息、判断任务归属，Worker负责具体执行。不同任务有独立的上下文，可以并行推进。
更关键的是，它试图让Agent主动找活——群聊里有人在讨论一个Bug，即使没人@AI，系统也能识别出来，在后台尝试复现、分析、甚至提PR。
这个方向如果跑通，Agent的竞争重心就彻底变了。
以前大家比的是“我的模型能不能把这道题做对”。
Super-Agent比的是“我的系统能不能在没人盯着的情况下，把一件持续好几天的事跑完”。
前者是模型能力，后者是调度能力、状态管理能力、异常处理能力。
而这些东西，光靠模型聪明是解决不了的。
你需要一个能让人往里塞东西的插件架构，需要事件流让每一步可观测，需要存算分离让长任务不丢状态。
DSH的“一切皆插件”设计，本质上就是在为这个方向铺路。
三、为什么“插件”成了新的竞争点？
我见过太多Agent框架在Demo阶段惊艳，一上生产就崩。
原因翻来覆去就那几个：工具调用没日志、长任务跑一半断了、换一个模型就要重写一遍调度逻辑。
插件架构解决的不是“能不能做”，是“能不能让别人帮你做”。
一个框架再强，官方团队就那么几个人。
你不可能预判到每个用户的使用场景。
有人要刷B站，有人要搜论文，有人要把女儿的画做成游戏。
这些东西官方做不了，但用户自己能做。
DSH的插件机制核心就两样东西：TypeScript插件 + cordis.yml。插件导出一个apply(ctx)函数，通过ctx注册工具、事件、定时器。
框架在插件卸载时会自动清理注册的资源。
这套设计不新鲜，但它足够简单——简单到“让DeepSeek自己写一个arXiv插件”这件事，一个普通开发者花一个下午就能跑通。
简单，才是生态能长起来的前提。
Codex、Cursor也在往插件生态走。
Codex的右侧工作区从“查看记录”升级成“多功能交互区”，方向就是让用户把“AI生成之后那5%的人工微调”也在Agent里完成。而补足这5%的能力，最可行的路径就是社区插件。
Agent的竞争，正在从“模型能力竞赛”变成“生态繁荣度竞赛”。
谁的插件多，谁就能覆盖更多长尾需求。谁的长尾需求覆盖得多，谁的用户就离不开。
四、但热闹背后，真正难的东西还没解决
评论区晒插件很热闹，Super-Agent的构想也很宏大。
但作为天天跟线上系统打交道的人，我得说一句不太中听的话：插件生态能解决“有没有”，解决不了“稳不稳”。
一个插件写得再好，它跟主框架的版本兼容吗？
一个Agent跑三天，中间模型换了两次、插件更新了一次，状态还在吗？
多Agent并行的时候，两个Worker同时改同一个文件，谁赢？
Super-Agent把“派活”这件事从人手里拿走了，但它同时把“保证活干对”的责任也拿走了。
这个责任，现在落在了调度系统、状态管理和异常处理上。
这些能力，比“让模型多调一个工具”难得多。
IDC的研究经理孙振亚在评论Super-Agent时说了一句话，我觉得很准：个人助理的竞争正在从任务式交付转向生态、记忆与管家式服务的综合比拼。
中国产品在记忆和任务能力上已经有基础，差距主要在服务生态壁垒。
生态壁垒，不是靠官方写一百个插件堆出来的，是靠一千个普通用户各写一个“我自己的那一下”长出来的。
五、写在最后
回到最初的问题：Agent的竞争，是不是已经换赛场了？
我的判断是：换了一半。
模型能力依然是地基，但它不再是区分度最大的那一层。
当一线模型的差距快速收窄，当“调通一个工具”变成几行代码的事，真正的分水岭就变成了：能不能让足够多的人，愿意在你的框架上造东西。
B站评论区里那些晒插件的开发者，那些写教程的UP主，那些把女儿的画做成游戏的普通人——他们不是在“用”一个Agent框架，他们是在参与定义这个框架能做什么。
DeepSeek的Super-Agent构想，赌的就是这件事：当Agent能自己派活、自己找活，人只需要做决策和验收，那框架的价值就不再是“我帮你做”，而是“我让你能教它做”。
至于这个赌局能不能赢，看评论区就知道了。
如果一年后，那里还在晒新东西，答案就是肯定的。
参考资源：
- DeepSeek Harness B站入驻视频：www.bilibili.com/video/BV1i2…
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。
Foreword
There's something interesting about the Bilibili technology zone recently.
DeepSeek Harness (hereinafter DSH) is officially stationed in Bilibili, sending a video, breaking 2.11 million on a 24-hour basis and increasing the account number by 400,000 a day.
But what really stopped me for a long time was the comment area.
It's not the kind of sofa "front row" water.
Some people are tanning their own plugins for DSH, others are discussing the details of a particular plugin, and others are asking “if there are 15 plugs on my side, there are more than me”.
Others have made their daughter's drawings into a platform jumping game that has received over 6,000 praises.
It's not an official organization, it's a group of people who make things spontaneously.
I've been doing Java for so many years, I've seen open-source project fires, I've seen frames go out, but a comment area of the Agent framework has become an exhibition of developers, and it's worth thinking about as much as any run.
It points to a question: Agent's competition is moving from “who's smarter model” to “who's more ecologically attractive”.
First, let's see what's going on in the comment area.
The DSH official video is actually about its “Creative Mode” — saying what it wants in a conversation, it writes the idea into a plugin and it works directly.
The video shows the countdown, the signing wall, the sandbox game, the 3D visualization of Transformer.
But the comment area is much more tanned than the official demonstration.
Somebody made a dsh cat plug.
A visualization of the near-modern history and Mau Chinese has been done, and the bottom row has sought links that are useful for study and study.
Mobile phone remote connectivity tools have been developed to optimize notifications and UI matching.
And the eye-eyed viewer found a line in the left-hand session list, "Build the AGI, make sure it's done." The screenshot came out many times.
These plugins have one thing in common: they are not “common capabilities”, they are “my needs”.
Professional developers write about infrastructure — memory, movement, remote connectivity. The ordinary lover wrote, "The moment I feel uncomfortable with every day."
Both sides are writing, both sides are throwing things into the same ecology.
Bilibili even has a complete curriculum ecology.
A long video, DeepSeek Harness, from the Introduction to the Field, was presented, from the bottom theory to the development of the plugin.
Someone did a desktop test, and step by step taught you how to make a PDF reader a DSH plugin.
There is also a field course on the Zhihu column, which shows "I made DeepSeek write himself an arXiv search plugin."
A framework can live without looking at the number of official documents and whether there is a willingness to write "second lesson" spontaneously.
What's Super-Agent in DeepSeek bet?
It's just another question about how far Agent can go: Who's gonna take care of these five when one of us runs the five?
DeepSeek, in a technical paper by Super-Agent dated 31 October from publish, put the pain in the face.
The majority of CodingAgent is currently centred on the Ssession, each of which only cares about its own share.
One can usually follow only three to five active sessions at the same time.
AI is becoming more capable and has become the “five AI project manager”.
Super-Agent's solution is Router-Worker's double architecture.
Router is responsible for receiving information and judging the attribution of the tasks, while Worker is responsible for their implementation. There are separate context missions, which can be pursued in parallel.
More importantly, it tried to get Agent to take the initiative - someone in the group chat was talking about a bug, and the system could identify it, even if it wasn't @AI, and try to re-emerge, analyze, and even refer to PR in the backstage.
If this direction runs, Agent's competitive focus changes completely.
You used to say, "Can my model do the right thing?"
Super-Agent compares to “Can my system run away from a thing that lasts for days without anyone looking at it”.
The former is a model capability and the latter is a movement control capability, a state management capability and an anomaly processing capability.
And these things can't be solved by modeling.
You need a plug-in structure that allows people to stuff things into it, an event stream that allows every step to be observed, and a separation that keeps the long mission alive.
DSH's "all-plug-in" design essentially paves the way in this direction.
Why has the “plug-in” become a new point of competition?
I've seen too many Agent frames pop up in the Demo stage, and it just goes down.
The reasons are those: tool call has no logs, long missions have been cut off and a model has to be rewritten.
The plug-in structure does not deal with “can we do it” but with “can someone else do it for you”.
With a stronger framework, the official team is just a few.
You can't predict the use of each user.
Bilibili is painted, papers are searched and girls' paintings are made into games.
It's not official, but the users can do it themselves.
The DSH Plugin Mechanism is based on two things: TypeScript Plugin + Cordis.yml. Plugin exports an apply (ctx) function through a ctx registration tool, event, timer.
The framework automatically clears registered resources when the plugin is offloaded.
It's not new, but it's simple enough -- it's simple enough to let DeepSeek write himself an arXiv plugin. In this case, an ordinary developer can run through one afternoon.
Simple is the prerequisite for ecology to grow.
Cordex and Cursor are also moving towards the ecology of the plugin.
The right-hand working area of Codex was upgraded from a “viewing records” to a “multifunctional interactive area” in the direction that the user completes the “manual fine-tuning of 5 per cent after AI was created” in Agent. The most viable path to complement these 5 per cent is the community plugin.
Agent's competition is moving from a “model capacity competition” to an “ecological prosperity competition”.
Anyone with more plugs can cover more long-tailed needs. Those whose long tails are much more covered are the users.
Four, but behind the scenes, the real hard stuff hasn't been solved.
The comment area is hot, and Super-Agent's vision is ambitious.
But as a person who deals with the system on the wire, I have to say something that's not very good: a plug-in ecology can solve “are there or not” and not “stable”.
A plugin is perfect. Is it compatible with the main frame version?
One Agent ran for three days, the intermediate model changed twice, the plugin was updated once, and is it still in shape?
The two Walkers changed the same file at the same time. Who won?
Super-Agent took the work away from the person, but at the same time took away the responsibility for the job.
This responsibility now falls on the movement control system, state management and unusual handling.
These capabilities are much more difficult than “leting the model adjust a tool”.
In his comment on Super-Agent, the research manager of IDC, Sun Zhenya, said that I think it is true that personal assistant competition is shifting from task-based delivery to integrated ecological, memory and housekeeper services.
China's products already have a foundation in memory and mission capabilities, and the gaps are mainly related to service ecological barriers.
The ecological barriers, not from the official writing of a hundred plugins, have grown from the writing of a “my own one” by 1,000 ordinary users.
V. At the end
Back to the original question: Has Agent's competition changed course?
My judgment is: half.
Model capabilities remain the foundation, but they are no longer the most differentiated layer.
When the gap between the first-line models narrows quickly, when the "one tool" becomes a few lines of code, the real watershed becomes: can enough people be willing to build on your framework?
The developers of sunplugs in the Bilibili comment area, the UP masters of the curriculum, the ordinary people who make their daughters' paintings into games — they're not “using” an Agent framework, they're participating in defining what this framework can do.
DeepSeek's Super-Agent vision, and that's the bet: when Agent is able to deliver and find his own life, people just have to make decisions and acceptance check, the value of the framework is no longer “I help you do it”, but “I let you teach him to do it”.
We'll see if this game wins.
If a year later, there's something fresh there, the answer is yes.
Reference resources:
- DeepSeek Harness Z59Z is posted on video: www.bilibili.com/video/BV1i2...
Auto-fetched and translated by AI Atlas. All rights belong to the original author.