I Rebuilt My Website With an AI Agent: An Honest Debrief

The site you’re reading was built, mostly, by an AI agent while I watched and steered. As someone who ships machine learning for a living, I wanted to see where the current generation of coding agents actually is: not in a demo, but on a real task with a real, slightly awkward requirement. Make it about me, using my actual background.

Here’s the honest debrief.

The hard part wasn’t the code

I assumed the interesting challenge would be the website. It wasn’t. The agent scaffolded a clean, fast site (semantic HTML, dark/light theming, a blog, project pages, even a generated résumé) quickly and without drama.

The genuinely hard part was getting my own data. I asked it to pull my details from LinkedIn, and that turned into a small odyssey: modern Chrome deliberately blocks automated access to your logged-in browser for security, a port it wanted was already taken by another app on my machine, and a shell-quoting gotcha killed a launch command outright. None of that is “AI is dumb.” It’s that real work is mostly integration and environment, and the messy edges are where the time goes. That maps almost exactly to my day job: the model is rarely the bottleneck; the plumbing around it is.

Where the agent was genuinely good

• The mechanical 80%. Boilerplate, config, consistent styling, wiring up a content system, writing a build that actually builds. Fast and reliable.
• Working from real artifacts. Once it could read my profile, it turned a résumé’s worth of raw text into structured pages without me babysitting the details.
• Explaining its own tradeoffs. It told me why it picked one framework over another, which made it easy to agree or overrule.

Where a human still had to steer

• Taste. Colors, tone, what goes on the homepage, how loud the hero should be. The agent had reasonable defaults, but “reasonable” and “me” aren’t the same thing.
• Truth. It cheerfully wrote a confident bullet point about my job that I had to flag as a placeholder, because it couldn’t know the specifics. Left unchecked, that’s how a nice website quietly starts lying about you.
• Deciding what matters. Which two projects to feature, what to leave off, what story the whole thing should tell. That’s judgment, and judgment is still mine.

The verdict

The “agentic AI will run everything” narrative and the “it’s all hype” narrative are both wrong, and building this was a small, concrete proof of the boring middle. An agent is a superb collaborator for the parts that are laborious but well-defined. It is not a replacement for taste, truth, or knowing what you actually want to say.

Which, honestly, is the most reassuring result I could have hoped for. The leverage is real. So is the need for a human who gives a damn about the details.

(Full disclosure: this post was drafted by that same agent, from a real session. I edited it. Steering, as ever, is the job.)