The brief was one sentence
I asked the agent running on my Hermes setup to critique this site and improve it. The old version had taste, but not much evidence. It also had the kind of technical SEO bug that looks harmless until every missing URL, including crawler files, returns the homepage with HTTP 200.
The agent audited the live desktop and mobile site, preserved a local baseline, searched for the missing source, then rebuilt it as dependency-light HTML and CSS. It kept the warm editorial character, but changed the job of the page: explain what I do, show enough operating proof to be credible, and make the next step obvious without inventing a client quote or a heroic metric.
It also added the boring things a generated mockup usually forgets: canonical URLs, social metadata, structured data, a sitemap, a real 404, security headers, self-hosted fonts, clean note routes, and a build script that publishes only intended public files.
“Looks good” was not the completion condition
The build had to pass a static regression checker, responsive browser tests, link and route checks, metadata validation, and Lighthouse. The agent found and fixed its own malformed JSON-LD before the site got anywhere near production. That is not glamorous, which is exactly why it matters.
The deployment path was deliberately dull: a private GitHub repository remained the canonical source; Cloudflare Pages built the main branch; a generated pages.dev address was verified before the public domain moved.
Authentication still belonged to me. The agent could navigate the control planes and diagnose why the first GitHub identity lacked authority, but it could not pretend a collaborator was the repository owner. I approved the one-time app access. The agent then caught Cloudflare steering us toward a Worker deployment and backed out before the wrong product shipped.
Cutover is where the demo ended
The new preview passed route, header, asset, crawler, metadata, desktop, mobile, 404, and Lighthouse checks. Only then did we move www.
The existing domain was attached to an older Pages project, so the agent first recorded the old deployment as a rollback target. It removed only the www binding, attached that hostname to the new Git-backed project, watched the brief 522 and 523 responses during propagation, and kept polling until the new title and headline appeared.
Production then passed 33 HTTP and parity checks plus six rendered browser routes. The apex domain still redirects to www. The previous site still exists on its old pages.dev address. Rollback is a URL and a binding change, not a prayer.
What the agent actually contributed
It did not independently develop taste, own the domain, or become legally responsible for the copy. It did the operational work around those decisions: inspect, build, test, trace authority, recover from the wrong path, mutate the approved systems, and read production back.
That is the version of autonomy I care about. Not a browser recording where the happy path happens once. A system that can move through files, Git, provider dashboards, DNS, and QA while preserving evidence and a way back.