Agentic engineering: how this site was built
I built this site with AI agents in four days, and kept building it the same way: plans before code, an agent for each task, my review on every pull request, and checks that run on every change.
- 4 daysfrom the first commit to the live site
- 95pull requests in the first three weeks
- 300+automated checks on every pull request
Problem
Coding agents write code fast, and that's exactly the risk: fast code that nobody planned, nobody checked and nobody can explain later. I wanted to find out what it takes to build a real site with agents and still own every decision, and the site itself was the test.
Approach
Interviews before plans, plans before code
Every piece of work starts with an interview: the agent asks me one question at a time, with a recommended answer for each, until we agree on everything. The answers become a document, then a plan of small steps, each with a line that says when it's done. The launch copy took 10 questions, the build plan 11, the site's assistant 12.
Agents build, I decide
Every step is marked as an agent's or mine. Agents write code, docs and tests. Purchases, DNS, accounts, secrets, legal text and final approval stay with me. The design came from Claude Design as an interactive prototype, and the code follows it.
Everything goes through a pull request
No agent pushes to main. Every change is a branch, a pull request, an AI code review and then my review.
Checks instead of trust
Every pull request runs formatting, linting, type checks, unit tests, browser tests on desktop and phone, accessibility checks in both themes, Lighthouse budgets, and a test that keeps the site from advertising anything. A pull request can't merge while one fails.
A log of what really happened
Every milestone, decision and problem goes into a log as it happens, including the mistakes.
What went wrong, and what caught it
- An agent committed generated files (
node_modules, build output) withgit add -A. It found the mistake itself while rebasing and rebuilt the branch. Since then, agents stage files by name and check the list before every commit. - A Lighthouse budget started failing pull requests that changed no site code. The agent read the reports, found that the score varied on CI runners, and came back with options instead of lowering the budget itself. I chose the new budget.
- A friend saw a certificate error on a company laptop. The agent's first theory, the certificate chain, was wrong. The real cause was the company's filter warning about a newly registered domain: registered 17 days earlier, exactly as the warning said.
- Writing the operations guide showed a security gap: a device removed from the dashboard stayed signed in for up to 12 hours. Sessions now end as soon as their device is removed, with a test.
Result
The site went from a first commit to live in four days, and has kept changing since: an AI assistant, a private dashboard, an email signature, 95 pull requests in the first three weeks. Every change now goes through the same checks: over 300 automated tests, more than 50 Storybook stories, and a 40-question test set for the assistant that I run by hand. What I learned is that the speed comes from the agents, but the quality comes from the structure around them.
