How Hekton Built Its Own Public Site

The Agentic Tekton is a public marketing site. It exists to make the factory look credible to people who will never read a line of its code. That is exactly the kind of project where it is tempting to cut a corner, ship the site fast, worry about the rigor later, since nobody outside the team is going to check whether the repo followed the usual rules.

I built it through the same pipeline as everything else in this factory anyway, classification, governance files, a human-apply gate on infrastructure, and the site that came out the other end was better for it, not despite it.

The actual problem: the site that argues for discipline was tempted to skip it

A site whose entire pitch is “here is how to build trustworthy software with agents” would have been a strange thing to build carelessly. But marketing sites are usually the first project people treat as exempt from their own standards, low stakes, no real users depending on it, why not just ask a model for a homepage and move on. I felt that pull. I ignored it.

What I did: ran it through the same classification and scaffolding as any other factory-output project

The first real action was classifying the project and handing the skeleton to the scaffolding agent, not writing a single line of homepage copy. One command produced the repo, a control plane, the governance files, and a first commit, before any model touched a design decision. The forty-odd scaffolded files cost zero model tokens and had zero chance of arriving subtly wrong, which meant the model’s attention was still fully available for the two things that actually needed judgment: the homepage design and the writing.

What got weird: the governance felt like the good part

I expected the classification and decision-logging overhead to feel like a tax on a project that was supposed to be quick. It didn’t. Classification settled the project’s paths, naming, and ownership before a single file existed, which meant there was no “we’ll tidy it up later” moment, the phrase that precedes tidying up later approximately never. Decisions got written down at the moment they were made, not reconstructed from memory afterward. Validation gates ran clean before anything counted as done.

What worked: the infrastructure gate held even on a project I was excited to ship

The site needs cloud infrastructure: a private storage bucket behind a CDN, a certificate, DNS. The Infrastructure Gremlin’s rule is simple: it plans, it never applies. It read a deployment manifest, wrote the Terraform, ran a read-only plan, classified the blast radius, and stopped. A human had to review that plan and decide, deliberately, whether to run the apply.

That gate is easy to respect in the abstract and tempting to wave through when you are excited to see your own site live. It held anyway, structurally, not because I was disciplined that day but because the agent was incapable of running the apply regardless of how I felt about it. The review caught the one thing worth catching, a blast-radius classifier flagging the plan RED over a missing best-practice setting rather than anything actually risky, confirmed it was a false positive, and only then let the apply go ahead. The site has been live at theagentictekton.com since.

The reusable pattern

Treat your own marketing site, your own internal tool, your own “this one doesn’t really matter” project the same way you’d treat a project a customer depends on. The classification and governance overhead is small and mostly automatic once the scaffolding exists. What you get back is a project you can still explain in six weeks, and an infrastructure change that never gets applied by accident just because you were excited to see it live.

What next

The site went live the same way it was built, one gate at a time, and it has been standing since. The site argues that governance is the feature, not the tax. The next real test of that claim is not the calm first build, it’s whatever change to this site someday needs to ship under real time pressure, since that is the moment corner-cutting actually gets tempting.