AI can write the code.

AI can write the code. Engineering is knowing what to test.

I’m currently working on a poker-vision system where the application needs to understand what is happening on the table — recognize cards, identify ranks, process board states, and translate visual information into usable application logic.

And this project is teaching me something important about AI-assisted development.

The difficult part isn't getting AI to change the code.

The difficult part is proving that the change actually works.

My current workflow looks more like:

Read the codebase → Understand the existing logic → Make the change → Run the test suite → Inspect the real output → Find inconsistencies → Fix → Test again.

For example, I’m currently testing multiple buyer frames against the recognition logic instead of assuming that one successful test means the system is correct.

That's a big difference.

A function can compile. A test can pass. The UI can render.

And the system can still be wrong.

That's why I don't want my AI workflow to be:

“Generate the code and move on.”

I want it to be:

“Make the change, challenge the result, and verify the behavior.”

This is becoming increasingly important as coding agents take on more implementation work. Modern agent workflows are already moving toward repository-wide investigation, automated tests, linting, code review, and validation rather than simply generating files. The GitHub Blog

For me, AI isn't replacing the development process.

It's changing where I spend my attention.

Less time typing every line.

More time understanding:

What should happen? What actually happened? Why are they different? And how do I prove the fix is correct?

That's where the interesting engineering work begins. 🔥

For developers building with AI agents: what do you trust more a passing test, the actual application behavior, or both?