A "Shrimp Farmer's" Insight: AI Agents Are Not That Wonderful

Written by: Haotian

As a “shrimp farmer,” after weeks of ongoing suffering and torment, I want to share some insights and reflections on shrimp farming for your reference:

1. Everyone knows that raising shrimp is for “improving efficiency,” but the truth is, at this stage, most ordinary people are “wasting time” on shrimp farming. Many unexpected troubles arise, such as Claude accounts being banned, API forwarding quotas being restricted, Openclaw suddenly upgrading and losing “memory,” etc. These issues consume most of the time, and the slight efficiency gains are far from proportional.

2. On Twitter, you see all kinds of posts selling AI anxiety—just look at them. Claims that a single command can make Claude control your entire computer, or that one prompt can make AI permanently take over your work and just relax—beyond that, it’s all nonsense. In reality, not only do they not relax, but often they code late into the night to fix a feature, fix bugs, and deal with new bugs. It’s hard to imagine how those who wouldn’t even spend five minutes actually operating the system can confidently shout about the arrival of AGI and AI overthrowing everything.

3. Raising lobsters can indeed realize many people’s dream of OPC (One Person Company), but the upper limit of large model capabilities is equal for everyone. However, people’s understanding and mastery of large models vary greatly. Don’t think everyone can become Peter Steinberger, Matt Schlicht, or Andrej Karpathy. The skills, framework design, iteration experience, ability level, and delivery results of top developers versus ordinary people are worlds apart.

4. Building your own AI OS through lobster farming is fundamentally about creating a personalized AI ecosystem. The more types of large models you connect in parallel, the more digital employees you manage, the more tasks you coordinate actively and passively, and the more nested skills and collaborative task combinations you have, the higher the chances of conflicts. The more demands like Cron jobs, real-time data scans, trading skills, etc., the greater the optimization challenge. Remember, lobster farming has never been about “model capability,” but about “engineering implementation and optimization.”

5. Equipping all digital employees with Opus 4.6 and Gemini 2.5 Flash Lite can achieve similar results. The former hires seasoned Wall Street elites, while the latter is more like enslaved labor from the slums. Maybe both achieve a function, but the costs and delivered results are incomparable. The truth is, the time you spend debugging and fixing bugs could be replaced by others simply by having enough “cash power.” Raising shrimp is really expensive, and the scary part is, most people understand this but still have to compromise and keep optimizing with inferior models.

6. Managing digital employees is like stacking LEGO blocks—more employees, more skills, more complex scenarios, and the risk of collapse in an instant increases. For some tasks, capability is enough; don’t be greedy or demand beyond your understanding. It’s recommended to put more effort into memory engineering, Git version control, removing model hallucinations, etc. Otherwise, many moments will push you to the brink of collapse. You might feel great after solving a cool requirement one second, only to see everything collapse the next, and want to cry. Don’t ask me how I know.

That’s all.