Is being acquired an inevitable fate? An exploration of the ultimate outcome of AI Agent startups triggered by CloudBot

The rapid rise and acquisition of CloudBot are not the end of Agent entrepreneurship but the starting point for moving from concept to implementation. Large models won't cover everything, and small and medium - sized teams don't need to be desperate.

Recently, the Silicon Valley and the domestic AI circle have been stirred up by two "lobsters": the open - source project CloudBot (OpenClaw, referred to as "lobster" by the community) has quickly gained popularity with its local execution and autonomous task closure. Subsequently, it was reported that the founder joined OpenAI, and the project received funding and was transferred to a foundation for operation. Almost at the same time, Manus, which focuses on general Agents, was acquired by Meta at a high price, becoming the two most eye - catching acquisitions in the AI application layer. For a moment, the industry consensus seemed to be simplified as: large models dominate everything, the space for independent Agent entrepreneurship is squeezed to the extreme, and being acquired is the only outcome.

The truth is far from that. AI Agents are not just the "skin" of large models, and small and medium - sized entrepreneurs don't have to survive only in the gaps between giants. This article takes CloudBot as a sample to analyze the real moat for Agent entrepreneurship, the possibility of replication in China, the survival logic under the pressure of large companies, and provides the implementable niche tracks and the final judgment in 2026.

01 The rapid rise of CloudBot: It's not just a shell, but a paradigm shift

First, clarify the core fact: CloudBot is not simply a chatbot that calls the large - model API. Its core value lies in five capabilities: local self - hosting, device execution rights, persistent memory, multi - model plug - and - play, and social entrance interaction. With a single user command, it can operate the browser, read and write files, process emails, and generate documents, achieving a leap from "giving answers" to "delivering results". This is also the fundamental reason why it quickly broke through on GitHub, boosted the sales of Mac Mini, and inspired global developers to follow up.

Its technical foundation is clear: the large model is responsible for reasoning, the MCP protocol is responsible for tool invocation, RAG is responsible for knowledge access, local memory is responsible for context continuation, and the interaction layer is grafted onto high - frequency scenarios such as Telegram and Feishu. This architecture is not bound to a single model and can access GPT, Claude, or domestic open - source models. In essence, it is an Agent gateway + execution engine, rather than the model itself.

So, does it have an independent moat? The answer is: There are product barriers in the short term, but no long - term technological monopoly barriers.

CloudBot's barriers come from four aspects: First, local privatization and privacy security, where data does not leave the device, meeting the sensitive needs of enterprises and individuals; Second, the interaction paradigm is extremely simple, allowing users to complete complex tasks in the chat box without opening a new app; Third, the community ecosystem and skill precipitation, with global developers contributing scripts and toolchains, forming a network effect; Fourth, execution stability, converting vague requirements into reproducible operation flows and solving the pain point of LLMs being "able to talk but unable to act".

However, these barriers are not insurmountable. Domestic developers can completely create products with equivalent functions within 1 - 3 months based on open - source frameworks and domestic large models. In fact, several desktop Agents similar to the "lobster" have been launched for internal testing, proving that the technological threshold for general - purpose Agents has been leveled by the open - source ecosystem. This also means that the entrepreneurial window period for pure general - purpose Agents is extremely short, and leading in product form cannot hold the ground in the long run.

02 Can it be done in China? Yes, but take a different path

There is no technological bottleneck for replicating CloudBot in China. Models such as Qianwen, Wenxin, DeepSeek, and Kimi have met the requirements of Agents in tool invocation, long - text processing, and logical reasoning; Low - code platforms such as Kouzibiancheng and Dify have significantly reduced the orchestration cost; In terms of private deployment, domestic system adaptation, and data compliance, local teams actually have more advantages.

However, directly creating a "Chinese version of the lobster" is doomed to fail. There are three reasons: First, large companies are quickly following up. Youdao has launched LobsterAI, and the Feishu, DingTalk, and WeChat ecosystems are all integrating desktop execution Agents. General - purpose entry - level products will be quickly covered by giants; Second, due to compliance and permission constraints, domestic supervision of device operation and automation permissions is stricter, and the path for pure C - end blockbusters is narrower; Third, commercialization is weak. The willingness to pay for general - purpose personal Agents is low, and it is difficult to support the team through donations and subscriptions.

The correct path for domestic entrepreneurs is not to replicate the entry point but to deeply cultivate scenarios. CloudBot has proven the viability of "AI that can do work", but it has not proven that "general - purpose working AI" can grow independently. Domestic teams should disassemble its technological capabilities and integrate them into the processes of vertical industries instead of competing for social entry points.

03 Can large models cover everything? No, because the boundaries are clear

A common misunderstanding: The stronger the capabilities of large models, the less necessary Agents are. This is a misunderstanding of the AI industrial chain.

The capability boundaries of large models are cognition and reasoning. They are not good at three things: Industry - specific in - depth know - how, process automation encapsulation, and end - side execution and permission control.

• General large models such as Alibaba's Qianwen and ByteDance's Doubao aim to cover all scenarios, so they have to compromise on vertical depth;

• Models like Kimi and DeepSeek have excellent application - side experiences but still stay at the level of "content generation + information integration" and do not have the ability to execute autonomously across software;

• The ecosystems of large companies pursue closed - loop operations and will give priority to adapting their own product matrices, providing insufficient support for third - party tools and niche industry processes.

The essence of AI Agents is the "hands and feet + industry knowledge + execution discipline" of large models. Large models provide the brain, while Agents provide action capabilities, industry rules, data security, and stable delivery. They are complementary, not substitutive.

This also answers the key question: Large companies have not compressed the entrepreneurial space to the extreme. Instead, they have opened up the underlying capabilities for free or at a low price, allowing entrepreneurial teams to focus on scenario - based, engineering, and commercialization aspects without having to develop large models. In the past, entrepreneurship involved three things: developing models, frameworks, and products. Now, only the latter two are needed, so the threshold has actually been lowered.

04 Is being acquired an inevitable outcome? Yes and no

Are Manus and CloudBot's successive integrations the fate of independent Agent entrepreneurship?

Looking at it in two categories, General entry - type Agents: The likely outcome is to be acquired or shut down. These products compete in terms of traffic, ecosystem, and capital. Giants can crush independent teams with ecosystem subsidies, and independent teams cannot resist in the long run. Acquisition is essentially about buying the team, product paradigm, and user habits, rather than irreplaceable technology.

Vertical - scenario Agents: They can grow independently. As long as they are rooted in the industry, have stable cash flow, and build data and process barriers, they have the possibility of independent listing. Overseas examples such as Harvey (legal) and Glean (enterprise search), as well as domestic Agents in government affairs, manufacturing, and cross - border e - commerce, have all verified the commercialization closed - loop.

So, being acquired is not the fate of the industry but the fate of general - purpose entrepreneurship. The outcome is determined the moment the track is chosen.

05 The real opportunity for small and medium - sized entrepreneurs: Abandon the general and focus on the vertical

In 2026, the keywords for AI Agent entrepreneurship are narrow door, in - depth cultivation, and delivery. Small and medium - sized teams should avoid the main battlefields of large companies and build barriers in the following four tracks:

Industry digital employees: Stable cash flow in the B - end

Targeting small and medium - sized enterprises, package Agents as "digital employees" to replace repetitive positions such as order - following, customer service, order review, reporting, and compliance inspection. For example, full - process Agents for cross - border e - commerce, Agents for quality inspection and energy consumption optimization in the manufacturing industry, and Agents for catering store operations. The core barriers are industry process templates + stable delivery + monthly subscriptions. Giants look down on small B - end scattered orders, which are the basic market for small and medium - sized teams.

Local privatization and compliance Agents: Meeting the security needs

In fields such as finance, government affairs, healthcare, and law, data cannot leave the domain. Develop vertical Agents with private deployment + permission control + audit trail to meet the requirements of security grading and data security. These projects have high unit prices and high renewal rates. The technological barriers lie not in the model but in compliance solutions and implementation services.

Lightweight automation tools: Compensating for the shortcomings of large companies

The ecosystems of large companies are closed and provide insufficient support for niche software and cross - platform collaboration. Develop small and beautiful automation Agents, such as cross - spreadsheet data synchronization, one - click distribution for self - media, converting design drafts to code, and batch resume processing, and adopt the plug - in and subscription model. Small teams are more agile and understand long - tail pain points better.

Agent low - code implementation services: Helping industries implement

Many traditional industries have needs but lack technological capabilities. Based on platforms such as Kouzibiancheng and Dify, customize industry Agents for customers, connect systems, debug prompts, and maintain knowledge bases, and earn implementation fees and annual fees. This is a light - asset, high - margin business that does not compete with large companies but becomes an ecological partner.

The common logic of these tracks is: Don't focus on the entry point, focus on capabilities; Don't compete for traffic, compete for delivery; Don't pursue the general, pursue the professional.

06 Final judgment: The Agent industry will move towards a three - tier structure

In the next 1 - 2 years, the AI Agent industry will form a stable three - tier structure:

• The bottom layer: Large - model and framework manufacturers, providing the brain and infrastructure, dominated by giants;

• The middle layer: Vertical - industry Agent manufacturers, deeply cultivating scenarios, having cash flow, and capable of independent development;

• The upper layer: Plug - in and tool developers, making lightweight innovations based on the ecosystem.

The general entry points will be integrated by giants, a group of small giants will emerge in vertical scenarios, and open - source and low - code will continuously reduce the innovation cost. Being acquired is not a failure but a reasonable exit for general - purpose teams; the real long - term value belongs to those who turn AI into industry productivity.

The rapid rise and acquisition of CloudBot are not the end of Agent entrepreneurship but the starting point for moving from concept to implementation. Large models won't cover everything, and small and medium - sized teams don't need to be desperate. Abandon the fantasy of "creating the next super - app", dive into an industry, solve a type of problem, and deliver stable value. This is the most solid way to survive in the AI era.

The significance of the "lobster" is not to tell us to create an entry point but to prove that AI that can do practical work is the real AI. And those who can do this in - depth will always have a stage.

This article is from the WeChat official account "Competition and Cooperation in Artificial Intelligence", author: Jinghe. Republished by 36Kr with permission.