How I build with AI agents in the loop

Most of my recent projects were built with an AI coding agent in the loop. That is a statement about workflow, not about authorship: I decide what gets built, how it is structured, and whether it is good enough to ship.

The loop

1. Specify first. Before any code, I write down the constraint, the interfaces and what "done" means — often as a short spec or a decision record.
2. Small, reviewable steps. The agent drafts one slice at a time. I read every diff; anything I can't explain doesn't merge.
3. Tests are the contract. Unit tests pin down logic, end-to-end tests pin down flows. For Bookodi Desktop that meant 115 unit tests, 21 end-to-end tests and CI runs that print through real system queues on macOS and Windows.
4. Verify on the real thing. Real printers, real phones, real browsers. Screenshots and traces beat "it compiles".
5. Write it down. Technical reviews and runbooks are part of the deliverable — they're how the next person (or agent) moves fast safely.

What it changes

• I can take on bigger surfaces alone — a cross-platform desktop app with five print transports, or a bilingual PWA with hundreds of tests.
• The bottleneck moves from typing to judgement: architecture, security boundaries and knowing which edge cases matter.

What it doesn't change

Security reviews, data-protection decisions and anything involving money or people get a human decision every time — the same rule I build into the agents I ship.