This site used to be a cyberpunk terminal: neon gradients, ASCII art, matrix rain. Fun to build, but it said “look what I can make” instead of “here is how I think”. The version you are reading was rebuilt with coding agents in the loop, and the process is a decent case study in what agent-assisted work actually looks like.
What the agents did well
- Cheap iterations. A full visual direction - layout, typography, copy - took minutes to implement. We went through several complete directions: an editorial portrait page, a Mediterranean flyer, a quiet minimal profile, a dark technical dossier, and finally this one. Trying them for real beat debating them in the abstract.
- Durable context. Every accepted or rejected direction went into a decision log in the repository. New sessions started from recorded decisions instead of re-discovering them, and rejected ideas stayed rejected.
- Self-checking. The site ships with a small test suite that locks approved copy, layout invariants, and even previously rejected phrasing. When the copy changed, the tests changed in the same commit, reviewed together.
What stayed human
- Every direction change started with me reacting to a rendered page. The agents never knew a layout was wrong until I said so. “This looks like a CV” is not something a test catches.
- The vetoes. The decision log is mostly a graveyard of my own rejections, which is exactly what makes it useful.
- The stopping point. Knowing when the page communicates what it should - authority without announcement, friendliness without noise - was never the agent’s call.
The lesson
The output was never the bottleneck; the judgment was. Agents made iteration nearly free, which moved all the value into the human loop around them: react, decide, record, repeat. That is the same shape I see in production agent work - the model provides the volume; the engineering around it makes the result trustworthy.
This is where I will keep notes on coding agents, production agents, and the judgment calls behind them.