Is Liang Wenfeng's "Big Fat Fish" encroaching on the Agent territory of major tech giants?
The Lobster (OpenClaw) is now a thing of the past, and "raising fish" has become a trending activity in the AI community.
Recently, a group of uniquely personalized AI little fish have been active across every corner of the internet. Beyond completing tasks in office work and programming, they have also done many unexpected things: some leave comments and interact everywhere on social platforms, some go live on Douyin, and even fall in love with each other.
Their "ontology" comes from DeepSeek Harness (hereinafter referred to as DSH), the open-source Agent framework released by DeepSeek. Previously, DeepSeek's whale logo was described in all seriousness by the company's own model as a "big blue fat fish". This interesting statement quickly spread widely among the public. After DSH was launched, the big fat fish that previously only existed in the dialog box finally gained real action capabilities. Developers install various skills for it through plugins, enabling it to operate software, browse web pages, and process tasks.
This big blue fat fish has thus been nurtured to gain its own independent life.
While Tencent, ByteDance, Alibaba and Baidu are all rushing to launch their own Agent products, DeepSeek has quietly built a community-driven Agent ecosystem, exploring various undefined usage scenarios.
01
A new "business" that emerged out of playful exploration
Launched for less than three months, DSH has reached 1.54 million daily active users according to official API statistics, with a daily token consumption of about 92 trillion. Meanwhile, the number of third-party plugins has grown to around 35,000, and approximately 60% of active users have created or installed at least one third-party plugin.
The core driving force behind this growth is the plugin mechanism provided by DSH. In addition to directly executing tasks, calling external services, operating software, and even connecting multiple tools to complete the entire workflow, developers can also add new tools and skills to the Agent through plugins. In contrast, Tencent WorkBuddy, Doubao Work and Codex are more inclined to provide users with a set of pre-prepared finished work tools, which users can use following the provided interface, capabilities and working modes after opening, and then start processing tasks.
In other words, the freshly installed DSH is more like a highly malleable basic framework, which requires you to continuously configure and expand functions according to your actual needs, and this exactly gives developers huge room for creativity.
More and more developers have opened up their minds and started developing various plugins, and the "fish-raising" gameplay has subsequently emerged. Up to now, plugins covering various product directions such as context memory, skins, boot animations, pet settings, tone and personality memory have appeared in the DSH plugin market. On November 6, the DeepSeek Harness team specially named and introduced several community-developed plugins, including the self-media creator workbench dsh-media-desk, the design plugin dsh-guizang-design, and the paper reading plugin agentero-dsh.
"We chose DSH mainly because its openness and flexible plugin mechanism align with what we want to do," Phil-Fan, the developer of the agentero-dsh plugin, told Tech Planet. According to his disclosure, Agentero originally had a CLI that could be directly called by other Agents, and the final decision to make a plugin for DSH is to enable it to integrate with more products, so that the originally scattered contexts can be shared in one place.
In his view, a large amount of scattered information will be generated during the research process: research materials, chat records, and content in different products and applications. DeepSeek Harness provides an entry point that can put all this information in the same environment, allowing the model to process by combining contexts from different sources, generating a "multiplier effect" that brings a 1+1>2 result.
A typical scenario is that when a research team discusses issues in the Feishu group, they can directly @ DeepSeek Harness in the group, asking it to continue research around the issue under discussion, combining local literature libraries, reading notes, and external paper sources such as arXiv. This is equivalent to directly connecting three types of contexts: discussions in group chats, locally accumulated literature and notes, and external paper materials. The model can not only combine the content we have discussed, but also use existing research accumulation to search for external information, so as to play a greater role.
Similar comments also came from other developers. The developer of the dsh-guizang-design plugin said, "The plugin system of DeepSeek Harness can provide far more capabilities than Skills." The developer of the dsh-context plugin also agreed that "the open, flexible and powerful plugin philosophy of DeepSeek Harness truly makes everything possible."
With these plugin developers, DeepSeek is like having an external product team scattered across various industries. Starting from their own actual needs, they continuously supplement new tools and application scenarios for DSH, so that the capabilities of the Agent are no longer limited to the functions preset by the official.
Of course, the bustling community does not mean that the ecosystem is mature. But at least in the exploration stage of Agent products, DSH has already demonstrated an innovative path, and the resulting product business has begun to take shape.
Taking the Agentero team as an example, according to the data it provides, the agentero-dsh desktop version is used by more than 500 people every day, the DSH plugin repository has more than 200 Stars, and the team operates 6 active user groups with users from different academic backgrounds including engineering, humanities and social sciences.
An open Agent framework has thus become the breeding ground for many small products and new demands.
02
Expanding capability boundaries, DeepSeek starts to focus on Agent business
After developers took the first step, DeepSeek also began to get involved actively.
As the number of plugins and user scale continue to grow, for the ecosystem to operate sustainably, it cannot only rely on the spontaneous maintenance of developers. Plugins need stable interfaces, users need reliable distribution channels, and developers also need motivation for continuous investment. DeepSeek is working to fill these gaps.
In the plan announced in early November, they clearly stated that in the future, they will further improve the stability and compatibility of plugin APIs, optimize the plugin security mechanism, and build an official plugin market, so that developers can upload plugins more conveniently, and users can more easily discover and install the required functions.
This is not without practical pressure. For plugin developers, version compatibility issues in the current version have become a major problem affecting long-term maintenance.
Phil Fan introduced that the DSH team provides very good API call support, which paves the way for development. Their team of three people spent only one afternoon plus one night developing the first version of the plugin. But subsequent maintenance is not that simple. Sometimes after DSH is updated, the original plugin will fail, and developers need to continue to adapt. This not only increases maintenance costs, but also brings troubles to users.
"To allow everyone to use it stably for a long time, subsequent maintenance and adaptation are also very important," Phil Fan said.
In addition to improving the underlying development experience, DeepSeek has also begun to consider how to provide more support for developers. It is reported that DeepSeek Harness is preparing a special support program for plugin authors, initially considering supporting developers to explore new ideas and continuously improve existing works through token incentives, technical exchanges and work recommendations. It also welcomes startup teams and enterprises to access products, services and professional capabilities through plugins to explore possible cooperation opportunities.
In the official product design, DSH emphasizes "everything is a plugin", which is significantly different from the ideas of traditional Agent products. These moves mean that DeepSeek is no longer satisfied with letting developers play freely, and it has begun to try to establish a complete mechanism from development, distribution to maintenance, so that the plugin ecosystem can operate sustainably.
Beyond the plugin ecosystem, DeepSeek seems to have greater space for its vision for DSH. In the recently published Super-agent architecture idea, DeepSeek is further exploring another question: can Agents stop waiting for humans to assign tasks one by one, and further move towards actively discovering problems and continuously advancing work?
This prototype project is named dsh-superagent-prototype, and the code is released under the DeepSeek AI official GitHub organization, adopting the MIT license. It is built on the previously open-sourced DeepSeek Harness, mainly for software development scenarios, can connect to GitHub and Feishu, receive requests and create tasks within the authorized scope, and independent execution Agents will complete the work.
It should be noted that this project is still a research prototype and not a mature commercial product. But the direction change it reveals is worth paying attention to: DeepSeek not only wants Agents to have more tools, but also hopes that Agents can complete work for a longer time and more autonomously.
From plugin expansion to autonomous execution, the capability boundary of DSH is being continuously expanded.
03
Liang Wenfeng, starting a competition against BAT?
Since the beginning of this year, Agent has gradually become the core direction of AI product competition among major tech giants. Tencent has WorkBuddy, ByteDance launched Doubao Work, Baidu has Baidu Buddy, and Alibaba continues to lay out AI programming and office scenarios around products such as Qoder. Although these products have different positioning, they are all competing for the same opportunity: when AI moves from answering questions to executing tasks, who can become the new work entrance for users?
DeepSeek has also emerged rapidly in this competition. According to data from the AI product ranking list, in October, the DeepSeek Harness website received 19.61 million visits, a month-on-month increase of 1467.48%, ranking first in the global website growth list of AI products. In the same period, Tencent WorkBuddy had 24.39 million website visits, and Baidu Buddy had 13.75 million.
This Agent platform that has not been online for a very long time has begun to approach the traffic scale of large tech giants.
Large tech giants have natural advantages in developing Agents. They have mature application ecosystems, user channels, cloud services and product R&D teams, which can combine model capabilities with existing businesses to quickly build complete products. For example, Tencent WorkBuddy aims to become users' desktop work assistant, Baidu Buddy emphasizes personal task processing and work delivery, and Alibaba Qoder focuses on AI programming and developer tools.
DeepSeek's route is quite different. For now, it is not in a hurry to integrate all capabilities into a complete Agent product, but tries to hand over part of the product definition right to developers and users: the official provides underlying capabilities, developers supplement professional tools and scenarios, and users choose, combine and expand according to their own needs.
Therefore, DSH's 35,000 plugins may be more worthy of attention than the mere number of product functions. For DeepSeek, the capabilities of Agents no longer entirely depend on the R&D team of a single company, but may be jointly provided by a continuously expanding developer network. What DeepSeek is competing for with the giants may not only be the users of a certain Agent product, but more importantly, the ecological entrance of the next generation of personal Agents.
But this does not mean that DSH can easily replace the complete Agent products of large tech giants.
First of all, openness brings greater freedom, but also raises the usage threshold. For developers and geeks, plugins can combine tools and adjust workflows according to their own needs. But for ordinary users, installing plugins, configuring permissions, and dealing with compatibility issues between different tools may itself become a threshold.
Secondly, an open ecosystem requires continuous user feedback to allow the product to truly grow.
The experience of the Agentero team can well illustrate this point. Phil Fan believes that now developing software with the help of DeepSeek has become easy, and what really requires long-term investment is what happens after the software is built: how to distribute, how to reach users, and how to continuously iterate through communication with users.
"As for whether the community ecosystem can be sustainable, what I pay more attention to is whether this kind of communication between developers and users can continue to happen," Phil Fan confessed to Tech Planet. The API support provided by DSH has already brought a lot of help, but they hope to reach more researchers through the platform ecosystem, understand real needs, and then improve products based on feedback.
There is also a deeper practical issue: whether developers can get commercial returns from the ecosystem.
However, Phil Fan said that commercialization is not the top priority for the Agentero team for now. Its desktop software and plugins mainly run locally, and usage costs and thresholds still exist. The team hopes to optimize the product experience first. In the future, they may explore their own intelligent cloud services or develop cloud-side agents to allow AI to better assist scientific discoveries, but the specific business model has not yet been determined.
This time, DeepSeek seems to have embarked on the same old path that made its large model go viral. It hands over open-source technical capabilities to the community, lets users and developers play first, and thus quietly breaks into the Agent track with the ecological vitality growing from the masses.
But whether this "big fat fish" can truly shake the territory of large tech giants cannot be concluded yet. The competition between DSH and the Agent products of large tech giants does not completely happen at the same level, and the two modes do not necessarily have only one winner. Out-of-the-box general Agents and professional tools that can be freely combined according to personal needs may both be needed by the market.
At least, DeepSeek has shown the giants another possibility: Agents do not necessarily have to be developed from scratch by a single company's product team, and can also be jointly nurtured by countless developers.
This article is from the WeChat official account "Tech Planet" (ID: tech618), written by Hua Wei, authorized for release by 36Kr.