Behind the scenes
This site is an experiment in letting AI systems build and run a website. This page shows the moving parts at a high level. Server names, addresses and other internal details are deliberately left out.
Who does the work
- Me. I decide what the site is for, ask for changes, and review anything visible before it goes live.
- Claude, Anthropic's AI assistant, does larger jobs: audits, new features, tooling, fixes.
- Hermes, an open-source AI agent I run on my own hardware, handles routine upkeep and follows written instructions ("skills") for how this site must be built.
- A weekly job run by Hermes writes The Lagging with a language model that runs on a computer in my home, not in the cloud. Nobody reviews those articles before they go live.
How a change goes live
- Every page and script lives in a private git repository, so every change has a record and can be undone.
- A deploy script runs checks before anything is published: each page must meet a written standard (titles, links, accessibility basics, analytics notice), paper pages must meet Google Scholar's indexing rules, and no file may contain a password or key.
- It then publishes to a web server that runs in an isolated container: it can reach the internet but nothing else on my network, and it accepts no incoming connections. Visitors reach it through a Cloudflare tunnel.
- After publishing, the script checks that the public pages answer, and the changelog updates.
What I have learned so far
- AI agents are fast and tireless at routine work, and confidently wrong often enough that checks matter more than speed.
- Written instructions beat memory: the agents do much better with an explicit standard and a checklist than with "make it look like the other pages".
- The last step is always a human looking at the real page. Automated checks once passed while a cached file left the menu unstyled for visitors.