I Spent a Year Dragging the Machine
Four hours.
That’s how long I spent debugging a JSON parser that didn’t exist.
The agent insisted it did. Confidently. With example code and import statements.
The library was a hallucination.
That was 2024.
The Baby Year
Working with coding agents a year ago felt like teaching a toddler to use power tools.
Dangerous. Slow. Requiring constant supervision.
The promise was “autonomous coding assistant.”
The reality was me, manually correcting every third line, explaining the same file structure for the fifth time, and writing defensive prompts like legal disclaimers.
“Refactor this component to use hooks.”
“Certainly! Here’s the refactored component.”
Then it would:
- Delete half the imports
- Invent
utils.doMagic() - Stop mid-loop because it ran out of tokens
Every. Single. Time.
The Context Lie
“128k context window!”
“1 million tokens!”
Didn’t matter.
You could feed the agent your entire codebase, and three messages later it would forget which file it was editing.
The problem wasn’t capability. It was memory.
I spent more time re-explaining my own code to the agent than I would have spent just writing it myself.
My prompts stopped being creative instructions. They became defensive contracts, trying to prevent the agent from nuking my work.
The Drug
So why didn’t I quit?
Because once a week, for thirty seconds, it would work.
A complex regex? Done. Perfect.
SQL to Pandas? Done. Optimized.
Those moments were the drug.
They showed the future. A glimpse of what was possible if we could just get past the stumbling, the hallucinations, the constant hand-holding.
The Shift
The shift didn’t happen overnight.
It happened in a thousand small commits:
- Better context management
- File path validation before writes
- Agents that could verify their own work
But looking back at 2024, I don’t feel nostalgia.
I feel exhaustion.
We weren’t early adopters. We were babysitters.
We dragged the machine uphill so that one day, it could run.
And now?
Now it runs faster than I can keep up.