I Spent $120 on AI Agents and Got 0 Users. Here's the Trap Nobody Warns You About.

I Spent $120 on AI Agents and Got 0 Users. Here's the Trap Nobody Warns You About.

At 11 p.m., you close the 17th page an agent generated.

You pay for three memberships. This month: $120. You text a friend: “I built a product.” The friend asks: “How many users?” You open the dashboard. 0.

You open the agent window again. “Optimize my landing page.” It says: “I'll look at visual hierarchy, information architecture, and conversion path.” You watch the cursor blink. You don't know what you're waiting for.

Ask the harder question. If everyone can use agents, and every top model is free, what makes you stand out?

AI has made execution cheap. Code, slides, images, copy, data analysis, small products. Skills that once took study and practice now take a prompt. The result looks decent.

So people think they are ahead. I use AI. I pay for the best plan. I installed the latest skill. I know prompting. I improved my workflow. Peel that away. When everyone can use agents, where is your edge?

That is the AI middle-income trap.

The catch-up effect

“Middle-income trap” came from economics. A country moves from low income to middle income on cheap labor and resources. Wages rise. Costs rise. The old growth model stops. Without innovation and industrial upgrading, the economy stalls.

Individuals are in a similar three-stage shift.

Stage one: AI is an advantage for a few. You can use agents to do what others cannot.

Stage two: AI becomes infrastructure. Prices fall. Tools get simpler. Using AI stops being rare.

Stage three: AI becomes default. Everyone uses it. The information gap closes. What separates people is what you do with AI, what you judge, and what you accumulate.

I took the college entrance exam in 2023. I chose computer science. I am not sure that was right. ChatGPT launched in late 2022. In fall 2023, GitHub Copilot was already there when I learned C++. Press Tab. Code appeared. I had spent hours on syntax, algorithms, debugging. Classmates could pass tests with AI.

I laughed at people who used AI for homework. Later I stopped laughing.

In 2024, I trained for ICPC. In 2025, Claude could solve many problems. In 2024, I made slides by hand. In 2026, a text-to-image model can produce a video from one line. In 2024, Cursor needed my supervision. In 2026, an ordinary user can tell an agent what to build, and it will build, test, and nearly ship it.

Score an outsider's knowledge out of 100. An outsider is 5. AI is 60 without guidance. Next month 65. The month after 70. A person needs years to move that far.

The lower bound of ordinary ability is rising fast. Most technical skills will keep losing value.

Models of the same intelligence get cheaper. What is expensive today becomes free tomorrow. Many jobs still use people because the math works. Once agents get cheap enough, those jobs are exposed.

The bigger risk: you mistake what your agent can do for what you can do.

The trap in practice

People spend $50 or $100 a month on AI subscriptions. They build toys. Ask about users or revenue: zero. That is not far from buying AI courses in 2023 and 2024.

You made a toy with one sentence. You improved your workflow. Paying for AI does not make you better. It can make you feel ahead.

That is the trap.

You get output now. You do not build capital for the long run.

You may say: I know prompting tricks. I install hot skills. I know every model's style. Is that not ability?

It is. But “using AI” splits into two kinds.

First: tool-bound tricks.

Which prompt works on this model. Which skill to install. How to tune a framework. These help now. They depend on tools being immature and users having different information.

Once a tool is useful, people make it simpler. Model companies make intent easier to read. Frameworks hide configuration. Media, communities, and course sellers spread what works. They chase their own interests. They turn expert tricks into default features.

Second: judgment and cognition.

Do you know the problem? What context must go to the agent? What counts as a good result? When should you trust or reject its answer? These look like “using AI.” They do not depend on one agent. The stronger the agent, the more these judgments matter.

If your edge is being better at the tool, it will shrink. Models, frameworks, and public experience all improve.

Move your edge from “using AI well” to “making decisions” and “knowing what is good.”

AI is rented. Capital stays.

AI ability is rented. What stays is capital that does not fall with model prices.

Capital can sit outside you or inside you.

External capital is older than agents. It can be users who understand you, use you, and give feedback. It can be an account with a following. It can be a workflow you keep improving.

Compounding matters. Past work lowers the cost of the next step. When execution is cheap, what you accumulated matters more than what you can make today.

Two people can make a product in a day with an agent. One still needs users. The other has users, an account, and brand recognition. AI closed the production gap. It did not close the distribution gap.

Apple released a foldable device. Developers ported the screen transition to MacBooks. The community made similar effects fast. Making it is not the same as getting people to use it. Apple has devices, an OS, distribution, and brand. An independent developer reaches a smaller group.

That gap is external capital.

External capital answers: where do you sit on the map?

Internal capital answers: how much can you lift with that position?

What remains scarce

In the long run: what is scarce when humans work with agents?

As agents improve, fewer things are human-only. Most productivity skills will lose value. Some will not.

First: judgment with consequences.

An agent can judge. It cannot take responsibility. If it fails, it apologizes. You pay.

Second: taste in service of people.

Make something people use. An agent can copy patterns. Whether it is good is decided by people. A user can say a design feels wrong. They may not say why. They may not know how to fix it.

Judgment is deciding in context.

Taste is knowing what is good.

They appear together. They train differently.

Training judgment

Cognitive science: the brain learns from the gap between expectation and result. Prediction before feedback helps correct bias.

Judgment loop: predict, see result, measure error, explain error, update model. The updated model predicts again. Reinforcement learning uses a similar shape.

Common failure: you hand a hard question to an agent. It gives an answer. You accept it. You feel smarter. There is no outlier. Your judgment does not update.

That is the trap in daily life. You pay. The agent works. You get the agent's output. You lose the rep.

Months later, the agent is stronger. You can do more. Your judgment has not grown. You depend on it more. People who started later catch up because tools got easier. Your edge dies slowly.

Before asking the agent, write your own call. It can be rough. Then discuss or test it.

Look for evidence that proves you wrong. Both you and the agent can be wrong. Use feedback from reality. Tie it to the call.

Turn “I was wrong” into “Why was I wrong?” Remembering the answer is weak. Understanding the error gives a rule you can use later.

Training taste

Taste grows by comparison.

Designer Iglika says taste comes from contact with many good products. You develop an eye. Thiago Costa, co-founder of Fey, pulls from art, architecture, film, and games. Not just his own field.

You need to admit some designs are better. Then you can ask why.

You also need to make things. As a Vibe Coder, you can use agents to make many versions.

Next time, do not ask “which is best?” Ask for several directions. Compare them as a user. Compare them with other products. You stop accepting the agent's first output. You stop blaming it for bad design. You train your taste.

Taste is a standard built from contact, comparison, and making, not “I like it.” It tells you what is better and why.

Less is more

Attention is limited. You cannot raise taste and judgment everywhere. Choose.

How to choose:

Importance. Does it relate to your field? Is it the bottleneck?

Long-term return. How good is the agent? How fast is it improving? What do peers know?

Transfer. Can you train it across tasks and reuse it?

Agents improve faster than any one person. If an agent is better than you at image generation, let it be average. Spend your attention elsewhere.

The danger is not losing skills. The danger is losing them without choosing.

Letting agents handle APIs, CSS, and slides is fine. That is progress. Letting taste and judgment fade without noticing is the risk.

Build a system

Agents are changing work. I noticed the trap pulling me toward mediocrity and dependence.

The response is not willpower. I am not good at “sticking with things.” I need a system that makes my own nature follow it.

Antifragile idea: a good system can use the agent era instead of being used by it. Change how you interact with agents. Create your own predictions, outliers, and comparisons. Track the project. Get better results. Build taste and judgment. That gives you a better chance in the long run.

Try these:

Prediction log. Before a decision, write your call, reasons, and expected result. Later compare with the agent and the result. Record the error.

Decision review. Ask why you judged it that way. What did you miss? What rule applies next time?

Multiple options. Do not take one answer. Ask for several. Compare, choose, explain.

Public output. Write your process, progress, and failures. Feedback builds external capital.

User feedback. Make things people use. Users force you to face consequences.

Selective degradation list. Name what you hand to AI and what you keep. APIs and CSS can go. Problem definition, aesthetic judgment, and responsibility stay.

Ending

You close the agent window. The screen goes dark. A blank document is on the desktop.

You write: “This feature will get 10 users.”

Tomorrow you open the dashboard. You look at the number.

On the next line: “Why 10? Why not 0? Why not 100?”

The cursor blinks. You type the first word.