How to grow a lobster: I spent three days testing an AI agent, and here are my findings about it

In early March 2026, a line of more than a thousand people lined up outside Tencent's headquarters in Shenzhen, China. "Every generation has its own eggs," they quipped, alluding to the way their parents choked in huge lines for free eggs, which in the PRC was at one time customary to give away as promos for stores, writes Chinese tech publication 36Kr. This time, people came for a free installation of the OpenClaw AI agent with a red lobster on its logo, which has become very popular in China lately. Of course, I couldn't pass by such a phenomenon. This text is written partly by an AI agent named Zwat. I share my personal experience.

OpenClaw is an open source AI agent created by Austrian programmer Peter Steinberger and released on GitHub in November 2025. An AI agent, technically speaking, is a harness that turns any large language model (Claude, GPT, DeepSeek - your choice) from a "conversational artist" into a real performer that can connect to mail, calendar, messenger, browser, file system. In fact, it can do the same things with a computer that a human can do: write letters, run reports, make restaurant reservations, trade stocks, write, test, and roll out code. Conventionally speaking, the big AI model is the "brain" and the AI agent is its hands.

And he does not expect a command from a person for every step - he breaks down the task into subtasks and crawls towards the goal inevitably, like a lobster on the seabed. Purposeful and a little scary. As Institutional Investor writes, in March 2026, OpenClaw will have 300 to 400 thousand users.

In China, it has become a national obsession called "growing lobster." According to Bloomberg, the wave of enthusiasm has captured all segments of society, from schoolchildren, students and office workers to retirees across the country.

Fu Sheng, founder of mobile game company Cheetah Mobile, while lying in bed with a broken leg during the long Spring Festival weekend in China, "raised" a lobster named Sanwan, who in 4 minutes sent Chinese New Year greetings to 611 contacts, wrote 14 articles to the company's official account, created a million-visit X thread, and took over some of the operational tasks. The story went viral.

At the same time on marketplaces there is an active trade of OpenClaw installation services from $4.9 to $499, and in San Francisco SetupClaw company charges $6 thousand for "premium installation" for top managers, writes 36Kr.

The big companies are not out of pocket either - whoever breeds lobsters at their place will pay the owners of AI models for their "brains". According to Bloomberg, the excitement around OpenClaw added more than $100 billion of capitalization to the Chinese technology sector in one week in March. As noted by Chinese state TV channel CGTN, in mid-March, Chinese companies processed 4.69 trillion tokens (units of information in AI) in one week, surpassing Americans.

"A chatbot uses only a few hundred tokens per conversation. A single active OpenClaw instance can consume dozens or even hundreds of times more tokens per day," Wired quoted Po Zhao, a technology analyst and founder of the Hello China Tech newsletter, as saying.

Now we understand why Tencent was giving away lobsters for free.

In addition, all major Chinese players launched their own "lobster clones": Tencent launched three at once - QClaw, WorkBuddy and LightHouse, Alibaba - CoPaw, Baidu - DuClaw, Zhipu - AutoClaw and ByteDance - ArkClaw. The real winner was MiniMax with agent MaxClaw: in mid-March, its shares soared more than 50% in two days and more than 500% since its IPO, surpassing Baidu - a company with $18.5 billion in revenue versus MiniMax's $79 million, Bloomberg wrote.

The Chinese state has not been left out - although it is looking at the lobster with some apprehension. Shenzhen has promised grants of up to 10 million yuan ($1.4 million) for "single-user companies" - startups where the entire staff is the founder plus his AI agents. Wuxi, a city near Shanghai, offered up to 5 million yuan ($730,000) for OpenClaw-based robotics breakthroughs.

At the same time, Beijing sent out warnings to banks and government agencies to restrict the use of OpenClaw. The agent requires broad access to data and can transmit information externally. According to Institutional Investor, there have been recorded cases of agents leaking financial data and cryptocurrency wallet keys, deleting emails and code libraries. "Give me a lobster and I'll turn the world upside down" sounds nice, until the lobster turns your email or database upside down as well.

In the West, there is no such consumer frenzy around lobsters as there is in China, but experts are well aware of the importance of the novelty. Nvidia CEO Jensen Huang called OpenClaw the largest, most popular and most successful open source project in the history of mankind. "It's definitely the next ChatGPT," he added.

Actually, Nvidia has already launched its NemoClaw platform for enterprise AI-agents. OpenAI CEO Sam Altman simply poached OpenClaw creator Peter Steinberger, and Anthropic has just announced the launch of its own AI agent that can remotely control a computer. However, only on paid plans and only for MacOS, but that's a start, as they say.

Somewhere I read an interesting idea: perhaps in the future, not a single person will be hired for a job, but a whole team of him and AI-agents trained and coached by him personally.

Before someone comes to me with a stern question, "Have you grown your lobster yet?", I decided to test OpenClaw.

I confess that I am an absolute zero in programming, I have never written a single line of code. At the same time, installation and configuration of OpenClaw, as it is written everywhere, is a rather complicated and troublesome affair, requiring the ability to write all sorts of fancy commands in the command line.For this reason, I decided to use the EasyClaw shell - essentially the same OpenClaw, but with a graphical interface "for dummies". Just what I needed!

Note: for security reasons, it is not recommended to install the OpenClaw AI agent on your work machine. It is better to use a separate test computer, or a cloud server (but it will cost money) or at least create a separate user on your computer without administrator rights. That's what I did, it's easy. The system is a regular Win-11 laptop on an Intel i7 with 16GB of RAM.

A separate task is to choose an AI that will become the "brain" of your agent. OpenClaw architecture allows you to connect almost any AI - from Claude and GPT to Chinese Qwen, Kimi or DeepSeek, as well as multiplatforms like OpenRouter, which gives access to dozens of AIs at once. Here, in general, everything depends on the goals - the most productive AIs are, of course, also the most expensive. Nevertheless, I bravely chose one of the luxury options - Claude Sonnet from Anthropic, which is now considered something of a gold standard for AI agents. Their flagship Opus is also powerful, but even more expensive. And Chinese models are many times cheaper (sometimes dozens of times cheaper), but are considered less suitable for large tasks.

From the tutorial videos it became clear that EasyClaw connects to its AI-brain through the API program interface, for which you need to buy and enter a special key, it's not difficult, it's done in two clicks.

After that I downloaded the distribution kit and started the installation. Here was the first surprise: the program never offered to enter an API key or showed me the fields where it could be done. It turned out that for even greater simplicity the authors of this lobster made it with already ready integration with a number of AI. To begin with, 200 free credits per day are given, which he spends on working with "brains" (you can choose more expensive or cheap), then you have to buy - a minimum of 4 thousand credits for $20.

But I had already bought the key! Well, I asked my lobster, whom I called Tzvat (צְבָת "claw" in Hebrew), in simple Russian words, what I could do about it, and he said: "No problem, I'll rewrite myself and connect to Anthropic directly, give me the key!". And he rewrote himself and connected.

Honestly, it looked like magic.

A little while later I complained to him that the money on Anthropic "burns" too fast and then he exclaimed: "But I have free credits, I'll switch to them immediately!". And he switched before I could say, "Stop!" But there was a nuance: there were only two free credits left at that moment, and while he was figuring out how to switch, the credits ran out and he froze. How I, swearing, with the help of Claude's chat prompts, rewrote Zwat back to Anthropic, I'll omit, it's too sad. But revived it and we moved on.

What I tested out of the skills. It can make digests of investment news on specified topics, say, at 9 am every day. At first glance, the digests are not bad, but some of the "news" turned out to be from February, but this can be customized if you work hard.

Integration with Telegram and other popular messengers is wildly convenient. You can control the work of your computer via Zwata right from your phone. For example, to ask to find a certain file (and not by exact name, but by approximate description like "send me the last of the digests") and send it to Telegram or mail. Yes, it manages mail, can sort or write letters. But it's better to get him a test mailbox on gmail, don't let him access your main mailbox, just in case.

I also asked him to study Lyft's Q4 financials, read the news and make an investment memorandum about the company, highlighting the unobvious pros and cons, and format it into a .pdf and put it in a specific folder. I'm not an investment expert, but I think the report turned out to be curious, such an exercise can give you some interesting ideas. The .pdf formatting was a bit lame in places, but that's probably fixable. Just in case, I remind you that before making any decisions, you should double-check all the facts for the AI.

Finally, I fed him the appropriate links and asked him to write a draft of the first half of this column. And in two versions - with a cheap Qwen "brain" and an expensive Claude Sonnet.

Predictably, it worked out much better with Sonnet. Which is somewhat comforting - if I'm going to be replaced by an AI soon, at least it will be an expensive one, not a cheap Chinese one! Speaking of money, it cost about $40 for three days of experiments with Zwat, which is quite a lot, but that's because of Claude. With Chinese Kimi or MiniMax, it would have been cheaper by about five times.Ideally, to optimize costs, you should use several AIs and switch between them depending on the complexity of the problem, BetterClaw advises.

To summarize, I can say that right now Zwat reminds me of a kind of friendly, not badly technically-skilled student intern, who wants to help, but sometimes makes mistakes due to lack of life experience and inability to foresee the consequences of his actions.

But, unlike a student, he doesn't sleep, he doesn't get sick, and he won't suddenly go to India for a year of enlightenment. I will continue to teach him about life, and we'll see what happens.

This article was AI-translated and verified by a human editor