Most of the code on my recent projects was typed by AI agents. That does not mean the projects built themselves. My job has shifted from typing to planning, directing and reviewing, and I am still the one who answers for what goes live.
1. Plan
Every project starts with a written plan: the pages, the data, the edge cases and what "done" means. The clearer the plan, the better the agents' work. A vague request produces confident but wrong code, so I spend real time here before anything is built.
2. Build
I use Claude Code, often with several agents working at once. They can read and change the repository, open a browser to see the result, and work on the server. One agent might build a page while another writes the API it needs.
This is where the speed comes from. Work that used to take me days of typing now takes hours, which leaves more time for the parts that need judgement.
3. Review
I read every change before it ships and test the site the way a customer would: on a phone, on a slow connection, with real data. When something is wrong, I send it back with a clear explanation, just as I would with a colleague.
What I always check myself
- Accessibility: keyboard use, screen readers and contrast. Generated code often looks right but fails here.
- Performance: page weight, image sizes and load time on a mid-range phone.
- Security: forms, uploads, authentication and anything that touches the server.
- Payments and customer data: never shipped without a careful manual review.
What it means for clients
Clients get working software sooner, and the time saved goes into the parts that make a site good: speed, accessibility and the details customers notice. The responsibility does not change. If something breaks, it is mine to fix.
This site was built the same way.