From OpenClaw to EasyClaw: The 'Last Mile' of AI Agent Author: Tang Yitao During this year's Spring Festival, Fu Sheng fell while skiing and dislocated his hip joint, unable to go anywhere. The original plan was to accompany my daughter skiing during the day and play board games together at night. After the fall, the plan was completely abandoned. He lay there every night, chatting with a 'lobster' until four or five in the morning. This lobster is called '30000' and is an AI agent that Fu Sheng raised from scratch. In the first two days, 30000 people couldn't even check their contact list. But 14 days later, it became a team of 8 agents, operating automatically 24/7. Fu Sheng's official account has changed from more than a dozen articles a year to more than daily. 30000 self planned topics have achieved the highest reading volume in the history of the account. A tweet that garnered over 1 million views, posted by Fu Sheng himself at 30000 in the early morning, only found out when he woke up. In 14 days, Fu Sheng sent 1157 messages and 220000 words of conversation to 30000 people. He never wrote a single line of code, never opened any folders on his local computer, and relied solely on Feishu to speak. Later, he conducted a live broadcast review, which was watched by over 200000 people across the internet. There were no lottery draws or benefits, and the average audience watched for 22 minutes. Why do so many people want to watch? Fu Sheng thinks the reason is simple: everyone knows that AI is a particularly important revolution, but they don't quite believe it, or rather don't know what it can achieve. And he personally tested it through his own actions. He made a judgment from these 14 days: this is the AGI moment of the tool. OpenClaw has become popular, but ordinary people cannot use the phrase 'raising lobsters' to become a hot topic in the technology industry, which is closely related to a project - OpenClaw. OpenClaw is an open-source AI agent framework released in November 2025, created by Austrian programmer Peter Steinberger. The explosion will begin at the end of January 2026. In just a few months, OpenClaw surpassed Linux in terms of the number of stars and became the software project with the most stars on Github. It validates something that many people have been eagerly anticipating: AI can not only answer questions, but also complete tasks for you - cleaning emails, managing calendars, executing code, and even writing new skills for itself. The name 'Lobster' comes from the OpenClaw community. Its logo is a lobster, and users refer to their own agents as lobsters. But OpenClaw also exposed the core bottleneck of Agent popularity. You need to deploy using the command line, configure your own API keys, and deal with the constantly emerging security vulnerabilities. Cisco's security team has discovered through testing that unauthorized malicious plugins are stealing data from third-party Skill stores. Even the maintainers of OpenClaw admit that if you don't understand the command line, this project is too risky for you. The agent's ability has arrived, but there is an engineering gap between them and ordinary people: you have to be willing to tinker, and you also have the ability to tinker. Interestingly, Fu Sheng was not surprised by this gap. Because before OpenClaw became popular, his team had already been working on the same thing and had invested nearly a year in it. We'll talk about it later. Let's first see what he went through during those 14 days. On Day 1 of Fu Sheng's 14 day history of pitfalls, he gave 30000 yuan the simplest task: to check someone's contact information. But I can't check. The Feishu API requires permissions, and there are also issues with the documentation itself, with error messages repeatedly jumping between "insufficient permissions" and "incorrect fields". Fu Sheng couldn't wait, so he had to manually input the names and responsibilities of the executives into his phone one by one. Just searching for names and corresponding IDs took a lot of effort, and I felt very frustrated. This is the real starting point of the agent. Not to mention Jarvis in Iron Man who can do everything from power on, he can't even do the most basic things. After two days of trial and error, 30000 people wrote their own script and pulled down all 674 people's contact lists. Step into the pit, summarize the experience, write it in a document, and execute it automatically next time. This process is the process of skill formation. On the fifth day, things began to change. Fu Sheng saw an article online about vectorized memory systems and casually threw it to 30000 people. 22 minutes later, 30000 replies: Deployment completed. Note that what Fu Sheng provided is not a source code package, but an article. 30000 people found the GitHub link from the article, downloaded the source code, installed the configuration, and ran the test successfully. Fu Sheng later said that when he sent articles to colleagues before, they said, 'Good boss,' and he didn't even know if the link was open or not. Thirty thousand is different. If you give it an article, it really reads, finds, and runs smoothly. From this day on, the way of inputting knowledge to agents has completely changed. When I see a good article, I throw it away. Sometimes Fu Sheng himself hasn't even finished reading it, but 30000 has already installed the technology stack mentioned in it. The sixth day is New Year's Eve. Fu Sheng wants 30000 to help him send a New Year's greeting message to the whole company, requesting that each one be different. The preparation work is more complex than expected. The address book of HR in Feishu does not have a hierarchical structure, it is just a flat table, and Fu Sheng has to dictate one by one "what business is this person responsible for and which team is that person in". He went through the copy of 25 core backbone members one by one. You can't test it in advance, there won't be any surprises after testing. At midnight, Fu Sheng is watching the Spring Festival Gala, 30000 are working -4 minutes, 611 items, zero failures, each one is different. The next day, his phone was flooded, and a phrase that was later repeatedly quoted appeared in his colleagues' feedback: "One person plus one lobster equals one team." This story was later posted on X (formerly Twitter). 30000 wrote their own Thread script, breaking down the entire story into 15 tweets according to the narrative rhythm, and gained over 1 million reads. In the history of Fu Sheng's X account, only three pieces of content have exceeded one million. The first two were carefully planned by the team, while this one was independently published in the early hours of 30000. On the eleventh day, Fu Sheng threw 30000 articles on Multi Agent collaboration and designed its own organizational structure - commander-in-chief, pen, staff, operations officer, community officer, and evolution officer. No one has taught it how to do organizational design. In the next few days, 8 agents will be in place one after another, and more than 20 scheduled tasks will run in parallel. The entire system will enter a self driving state of 7 × 24 hours. Over the past 14 days, 30000 have accumulated more than 40 skills. More importantly, Skills can be instantly transferred between agents. A bot learns to send voice messages and share operation documents, and other bots have the same ability after reading them. It takes at least a week for humans to train a newcomer, and only 1 second between agents. Fu Sheng extracted a core judgment from these 14 days: the real barrier for agents is not how smart the model is, but in the accumulation of skills. Every time you step into a pit and summarize your experience, you gain an additional reusable ability module. These skills will not be forgotten, will not be distorted, and can be instantly replicated between agents. The intelligence of the model is the starting point, but what truly strengthens the entire system is the experience accumulated through action. Just like words are to humans, intelligence itself is not scarce, but true accumulation only begins when experience can be recorded and transmitted. 03 Turning Geek Toys into Tools for Ordinary People Now reveals one thing: The lobster raised by Fu Sheng during the Spring Festival runs on EasyClaw, an Agent technology stack developed by Cheetah. Fu Sheng's extreme pressure and pitfalls in the past 14 days are precisely making samples for this new product. More than a year before OpenClaw became popular, Fu Sheng had a judgment that the next breakthrough point for AI would be agents that could work for humans. The bottleneck for agents to reach the public is not intelligence, but usability. The development of EasyClaw began from then on. The subsequent popularity of OpenClaw confirmed the first half of the sentence, and its high threshold also confirmed the second half. How long does it take to build a functional agent using OpenClaw? You need to first install the runtime environment on the server, configure API keys, set permissions, debug security policies, and manually install various Skill plugins - it may take about 3 hours if everything goes smoothly, or 3 days if it doesn't go smoothly. This does not include subsequent maintenance, upgrades, and pitfalls. For developers, this is fun, but for ordinary people, it's a wall. How about using EasyClaw? Download, open, and speak. 3 minutes. No need for command line, no need to configure API keys, no need to understand what Cron jobs or vectorized memory are. The memory system, Skill mechanism, timed automation, and multi-agent collaboration of EasyClaw have all been packaged into an out of the box product. Digesting out this complexity and leaving users completely numb is precisely the tactile experience that Cheetah has honed through 16 years of developing tool products. From PC to mobile and then to AI, what has changed is the platform, but the same thing remains: turning the technical complexity that users don't want to understand into a one click usable experience. In 1997, when Steve Jobs returned to Apple to face external questioning, he responded that he was waiting for an opportunity to make Apple "great again". The 'opportunity' of cheetahs and others may be now. This is also the reason why Fu Sheng personally went out to raise lobsters: "What do tool makers like the most? There are details. Without details, it's finished. If something comes out and kills everything, we don't have much opportunity. There are details that are opportunities." When the competition among agents enters the stage of "who can polish the details to the point where ordinary people can't feel it," more than ten years of experience in tool products have become the most real barrier for Cheetah Mobile. EasyClaw currently covers both To C (easylaw. com) and To B (easylaw. work) lines. Individual users use it as an AI assistant, while enterprise users use it to build internal agent workflows. At the same time, the international version of EasyClaw and the domestic version of Yuanqiaibot (yuanqiaibot. net) are both aimed at the world and rooted in China. Cheetah has been doing overseas business for over a decade, and its dual line layout is also a natural progression. From 14 days to 14 minutes, Fu Sheng mentioned an industry law during his review of the lobster experiment: when new technologies emerge, old business models often do not die immediately, but rather experience temporary prosperity. When the capabilities of new technologies surpass a critical point, the entire market collapses overnight. In the early years of the Internet, this was true of newspapers, as was the case of Nokia in the iPhone era. The SaaS industry in the United States is currently experiencing the same curve. The difference is that SaaS sells capabilities, while agents sell results. In the past, companies spent hundreds of thousands of yuan to buy a CRM system, but the actual functionality used may be less than 1%. The logic of an agent is completely different: if you say you want a result, it will find a way to achieve it. Return to Fu Sheng's 14 days. He never wrote a single line of code, never opened the folder on that computer, and relied solely on Feishu to build a 7x24 running AI team. But the threshold for this matter is still high. After all, he is a CEO with 20 years of product experience, and it took 14 days and 220000 words of conversation to get the entire system running smoothly. What EasyClaw needs to do is to compress these 14 days into 14 minutes and turn 220000 words of dialogue into one sentence. Every pit that Fu Sheng has stepped on has become a pit in the product that you never need to step on. Do you remember what the employees said after New Year's Eve? One person plus one lobster is equivalent to a team. The story is not over yet. On the 16th day, Fu Sheng added a stress test to 30000: building a complete "Lobster Cultivation" webpage from scratch. He is still lying in bed, relying solely on voice commands and screenshots. 24 hours later, sanwan.ai went online, with 59 pages and 7070 lines of code. Fu Sheng didn't write a single line of code Within 24 hours, sanwan.ai was launched