The new duploflow.com was built by AI agents. Agents proposed the changes, automated checks verified each one, and people set the direction and approved what went live. This update explains how that worked and what the build numbers show.
How the work was done
Each change began as a small, specific task, with its acceptance criteria written down before any work started. An agent proposed the change. Automated checks then ran against it: formatting, linting, tests, dependency and licence checks, a full build, link checks, and checks of the built pages themselves. A change that failed a check went back for another attempt, with the failure in front of it. Only changes that passed were merged.
People set the direction throughout. They decided what the site should say and how it should look, and they approved what reached the live site. Automated checks can tell us a page is broken or inaccessible. They cannot decide what Duploflow should say about itself.
The numbers
- 21 tasks went from the first commit to launch.
- 36 attempts in total. 13 tasks passed their checks on the first attempt, and the rest needed two, three or four attempts.
- 103 CI runs from the first commit to the production cutover. Not every run passed. The failing runs are how defects were caught before they reached visitors.
- About 11 minutes for a full pipeline run at the median, from push to deployed build.
- 100 in all four Lighthouse categories (performance, accessibility, best practices and SEO) on the home page, the newsroom, an article and the privacy page. Each score is the median of three runs.
- Zero axe accessibility violations in both light and dark themes, and zero errors from the HTML validator.
The launch itself was automated too. The new site was checked before and after the switch, so a problem with the switch would have been caught rather than left for a visitor to find.
What comes next
We will keep using the newsroom to report on how the company works and what it builds. Press enquiries go to press@duploflow.com.