Shipping with AI in the loop
How I use coding agents day to day without losing track of what the code actually does.
By Emmanuel Olowu
AI agents write a meaningful share of my code now. They have not made me less careful. They have changed where the care goes.
Plans before prompts
For anything bigger than a small fix, I write a short plan first: the goal, the files involved, what done looks like. The agent works from the plan, and the plan becomes documentation.
Small, verifiable steps
I ask for one change at a time and run the type checker, tests and build after each. An agent that can verify its own work is far more useful than one that writes a lot of code quickly.
I still read every diff
Generated code is still my code. If I cannot explain a change in review, it does not ship.
Where it shines
Boilerplate, migrations between library versions, test scaffolding and first drafts of admin screens. Where it needs the most supervision: security boundaries, money and anything involving time zones.