DEV Community

Hossam Assadallah
Hossam Assadallah

Posted on

What Training an AI Agent Taught Me About Letting Go of Control

his is the first English piece from "Psychology of the Keyboard," a series I write about the mental side of being a programmer — the part nobody puts in the documentation.

I was training the AI agent behind Amin, the product I've been building, and I had the exact reply mapped out in my head for every question it should get. I wrote the prompt, reviewed it, tuned it, ran the test. The reply came back — not wrong, exactly, just not phrased the way I'd already decided it should be. I went back, adjusted, tried again. A third reply came out, different from the first two. And I felt something strange. Not frustration at the model's performance — a feeling that I was losing control over something I had built from scratch.

I paused the training. Not because the model was wrong, but because I couldn't tolerate a system I'd built with my own hands producing responses I didn't control 100%. That made me think — this wasn't the first time I'd felt this. It shows up in everything I build, from the database layer to the UI: I want every detail under my control, completely.

And that instinct makes perfect sense in our profession. We're trained to believe that one point outside our control can break the whole system, so we've been conditioned to treat total control as the only real safety. The problem is that mindset doesn't stay inside the code. It follows you out. You catch yourself wanting to control an employee's reaction, a client's decision, even your own kid's mood mid-conversation — because your brain has learned that "no full control means danger."

But an agent, like any AI system, is built on probability, not fixed rules. It's not supposed to give you the exact same wording every time — that's not a flaw, that's its nature. The real lesson wasn't figuring out how to make the output 100% consistent. The lesson was learning to separate "controlling the general direction" from "controlling every word." I needed to tune the framework and the rules, not micromanage every small detail inside it.

Once I let the agent take up more space inside the framework I'd set, the responses got more natural and closer to each customer's actual tone — not because I'd gotten less precise, but because I'd gotten confident that the framework itself was enough.

Have you ever caught yourself trying to control something that was never meant to be fully controlled, at work or outside it? How did you know it was time to let go?

Top comments (0)