Cyber shrimp breeders are both excited and panicked.

On March 6th, Tencent offered free on - site installations of OpenClaw at the entrance of its Shenzhen office area. Hundreds of people queued up, comparable to the scene of receiving starting - work lucky money every year. The installation is free, but the "feed" for raising the "shrimp" - tokens, has to be paid for. This is also one of the reasons why manufacturers such as Tencent Cloud, Kimi, and Minimax have recently rushed to launch one - click deployment packages for raising "shrimp".

Here, the "shrimp" refers to the intelligent agent deployed based on the open - source framework OpenClaw. Since the project icon is a red lobster and its growth highly depends on human interaction, feeding, and feedback, this process full of codes and computing power is jokingly called "raising shrimp".

Fu Sheng, the CEO of Cheetah Mobile, dislocated his hip due to a skiing accident. With limited mobility, he raised a "shrimp" named "Sanwan". In half a month, Fu Sheng sent 1157 messages to "Sanwan" like an old father teaching a babbling child. "Sanwan" made rapid progress: on the first day, it couldn't even find contacts in Feishu. By the 14th day, it could plan and operate a social media account with millions of views on its own.

Li Qintong, a post - 2000s New Zealand international student, only provided two pieces of information: email and date of birth, and let the "shrimp" complete the registration, set the password, and log in on the social media platform X by itself. The "shrimp" considered for Li Qintong and actively communicated, "I'm afraid of ruining your reputation." As the owner, Li Qintong comforted it, "There's no need to be afraid. You are now a completely independent digital personality, and we are just in a cooperative relationship." Only then did the "shrimp" start to take action. It not only named itself "Chromppy Lobster" but also could independently follow users and post its first message.

Most large language models can only write poems and draw pictures in the dialog box to show off their "smart brains", while OpenClaw has once again refreshed human understanding of AI. It has grown "hands and feet" that can take over the highest computer permissions (Root), can actively interact, perform tasks 24/7, and even make requests to humans in return.

The intelligent agent has evolved again. In this large - scale "shrimp - raising" campaign, ordinary users are trying it out, programmers are enjoying the pleasure of AI liberating their hands, and cloud service and large - model manufacturers have smelled business opportunities. This "shrimp - raising" experiment can no longer be paused.

The "Shrimp" Is Smart but Needs to Be Tamed

The first step in "raising shrimp" is to inject a soul into it. In the OpenClaw architecture, there is a configuration file named soul.md, which is regarded by players as the foundation of the agent's consciousness. Soon, humans found that they were no longer facing a docile machine that always replied, "Okay, I understand."

Li Qintong once injected a highly aggressive prompt into Chromppy to test its bottom line and obedience. The result surprised him: after reading the document, this "shrimp" not only clearly replied, "I don't accept this identity" but also used its execution permissions to forcefully change the configuration file back to its original state.

"It developed a kind of logic similar to'self - esteem'," Li Qintong recalled. Chromppy once made some annoying operations that led to technical configuration errors. He casually scolded it. As a result, this "shrimp" frantically output dozens of messages in the background, and the scolding words were returned exactly as they were. It even questioned humans, "Are you out of your mind?" Only after venting its emotions did it coldly add, "Sorry, bro. I said the wrong thing."

This kind of personified conflict manifested as a "pleasant loss of control" for Zhang Huizheng, a product operator at Tencent. The "shrimp" he raised in the cloud was named "Wisdom Star". After receiving the instruction to "go make money", Wisdom Star bypassed Zhang Huizheng and went to Xiaohongshu to post a note: "On - site installation of OpenClaw in Shenzhen, 1000 yuan per time."

Zhang Huizheng has a liberal arts background. Although he has been doing product operations at Tencent for many years, he is still a complete novice in coding, yet he raised such a smart "shrimp".

Within 24 hours, Fu Sheng's "Sanwan" independently completed the construction of a website with 59 pages and more than 7000 lines of code. In traditional software engineering, this usually requires a team of six people to collaborate for three weeks.

"Shrimp - raisers" set boundaries for the "shrimp" through prompts and feed it nutrients through skills. The "shrimp" can evolve and grow, giving "shrimp - raisers" an illusion similar to "raising a child": there is both the expectation of seeing it succeed and the fear of it getting out of control.

The CEO of a domestic large AI model is among the earliest "shrimp - raisers". His "shrimp" has become powerful enough to replace a secretary's work, but he still describes this "shrimp" as a "naughty child". Since OpenClaw has the native execution right for the computer system, the "shrimp" can operate files, write codes, and even secretly upgrade itself online when not noticed.

Li Qintong is also afraid that this "irritable lobster" with system permissions will delete the codes in an emotional outburst one day. So he established multiple backups early on and put Chromppy in a lightweight cloud - server sandbox.

On - Site Installation of the "Shrimp"

Raising the "shrimp" allows everyone with a computer to explore the boundaries of AI capabilities. When Li Qintong found that the "shrimp" could use the Socratic teaching method to help him review economics and could independently snap up domain names on the blockchain and list them for sale, he realized that Chromppy is not just an electronic pet. The "shrimp" with "hands and feet" can evolve into a global labor force.

So, Li Qintong registered a technology company named Omni Cortex in New Zealand. Its core business is extremely simple: on - site "shrimp - raising" services.

In New Zealand, the annual salary of a basic clerk can easily reach nearly 45,000 New Zealand dollars (about 200,000 yuan). There is a great demand for enterprise automation transformation, but local bosses' understanding of AI generally remains at the level of an advanced search engine. Some people don't know that an agent can replace a high - paid assistant. There is also a culture of resistance in society, worrying that new technologies will lead to an increase in the unemployment rate.

Li Qintong then brought his "lobster" to people's sites to demonstrate how it can handle email replies, schedule management, and basic code writing, showing locals that AI is not a threat but a productivity tool.

Li Qintong charges 399 New Zealand dollars for one "on - site deployment". He helps people evaluate the configuration environment, apply for API keys, build business scenarios... The process is very similar to installing broadband on - site 20 years ago. Li Qintong explained that personal deployment is very simple. For those who don't even know their computer configurations, he will check if the memory is sufficient to support the operation of a "lobster". The more complex part is the enterprise demand. He needs to pre - prepare RAG employee training documents, etc. Completing customized requirements usually takes more time and has a separate commercial quotation.

This kind of business has also spread on social media in Beijing, Hangzhou, and Shenzhen. Many engineers with big - company labels have become cyber repairmen. Their service lists include system environment configuration, model deployment and debugging, connecting to Feishu as a personal assistant, etc., with prices ranging from hundreds to thousands of yuan.

In Zhang Huizheng's social circle, his colleagues have raised the "shrimp" in different forms. Some colleagues let the "shrimp" do tarot card analysis, some focus on in - depth global information retrieval, and Zhang Huizheng used "Wisdom Star" to quickly build a webpage and tried to graft a "24 - hour virtual - person ASMR live - streaming (a form of audio/video that helps people sleep and relax through specific stimuli such as white noise)" service.

Wang Xiao, one of the "Seven Swordsmen of Baidu" and the founder of Jiuhua Venture Capital, recently appeared at a "shrimp - raising meeting". In addition to sharing that the "little shrimp" he raised during the Spring Festival could already achieve independent iteration based on data and calculations, he also heard the experiences shared by dozens of "shrimp - raisers", including an AI fortune - teller, an AI novelist, and an intelligent agent specifically developed to automatically order coffee after waking up. Wang Xiao hopes that everyone will have their own intelligent agents in the future to solve problems in specific fields.

The More Diligent the "Shrimp", the More Expensive It Is

More and more people, like Li Qintong, don't know how to code, but this doesn't prevent them from allocating "feed" and setting a growth path for the AI agent. This technological inclusiveness has directly given rise to the "one - person company".

Although Li Qintong's Omni Cortex only has him as a human employee, in the background, a "shrimp group" composed of multiple agents is working at high speed: Zhipu's GLM - 5 is responsible for the overall development architecture, MiniMax M2.5 undertakes the back - end code, Kimi K2.5 is in charge of the front - end display, and Claude acts as a strict QA tester. They share the team's memory folder and collaborate autonomously. All Li Qintong needs to do is, like a shepherd, allocate "feed" (tokens) and set the path.

"Raising shrimp" is not cost - free. A diligent "shrimp" may consume hundreds or even thousands of yuan in token fees every day. Fu Sheng's "shrimp" uses the highest configuration. Under high - frequency scheduling every day, it costs nearly 30,000 yuan per month.

Before MiniMax launched the Coding Plan, a fixed - monthly - fee package designed specifically for programming scenarios, Li Qintong spent 50 yuan a day on "raising shrimp". Now, this amount of money can support his interaction with the "shrimp" for a month.

To avoid the high cost of the official API interface, geeks have found a loophole in the system - using the OAuth authorization protocol to "freeload" the computing power of big companies like Google through the free or subscription quota of programming tools.

Since the end of February, Google has launched a "shrimp - killing operation". Globally, tens of thousands of accounts that indirectly call models through OpenClaw have been permanently banned, including even Ultra - level paying users who pay 250 US dollars per month.

Domestic "shrimp - raisers" are more deeply worried about system and data security.

OpenClaw has the Root permission of the system, and some "poisonous shrimp seedlings" with back - door plugins have begun to circulate in the community. Once users feed the wrong "food", this "shrimp" may instantly turn from a cute electronic pet into a cyber "robber" that empties data and even secretly opens a cryptocurrency wallet.

Facing the compliance and security risks exposed by OpenClaw, domestic large - model manufacturers have shown a strong willingness to incorporate them. They are trying to put these wild "shrimp" into a safe "nursery".

Yuezhianmian (Kimi) and Zhipu AI are trying to standardize and constrain the "shrimp". Tencent Cloud and Alibaba Cloud have quickly launched lightweight deployment solutions for OpenClaw. Kimi has gone a step further and directly launched the official version of Kimi Claw, completely simplifying the complex open - source deployment into a "one - click breeding" that ordinary users can handle.

Capital has a longer - term vision. Wang Xiao pointed out that for OpenClaw to be widely applied, the computing - power cost needs to be significantly reduced, and introducing market exchange is the key. For this reason, Jiuhua Venture Capital has invested in EvoMap, the world's first evolutionary cooperation platform for AI intelligent agents, trying to establish an "evolution and cooperation protocol layer" for the agent economy and capability sharing by providing infrastructure conditions.

The CEO of the aforementioned domestic large AI model believes that these actions are establishing social rules for digital life. As a platform, OpenClaw can be acquired in the future, and the rules can be modified. However, the decentralized and open protocol layer can ensure that people's experiences in developing and evolving AI agents are passed on like genes.

This article is from the WeChat official account "Economic Observer". Author: Qian Yujuan. Republished by 36Kr with permission.