I used Gemini in Google search to finish a complex coding task for me, porting a library to a Godot GDExtension.

The AI gets nothing right on the first try. The workflow is:

- The AI generates something that doesn't work but is sort of right.
- I cut + paste the error message.
- The AI fixes that bug, but now there's a new bug.
- Repeat the cycle 20 times, eventually it finally works.

The AI doesn't seem to really actually understand the thing it is implementing. What's it's doing is generating slop, then it looks at the error message to fix that specific error.

However, now that I see the end results, I actually do understand what it generated, so now I can fix or improve it on my own.

That's what happens under the hood of the coding agents. It's doing a cycle of code, compile, fix error, repeat. It runs the cycle until it gets something that works.