Why I'm building AI agents in public

What does “building in public” mean here?

Every agent I build gets a public project page with the same fixed sections: what it does, why, the architecture, the stack, measured cost and latency, evaluation results, and a list of what broke. The code goes on GitHub. The numbers on the page come from run logs, not from memory. The first entry — an agent wired to real tools through the Model Context Protocol — is already up, with its full video walkthrough.

Why not keep a private notebook?

Private notes forgive vague claims. “It works pretty well” survives in a notebook; it does not survive a page that shows an eval number next to the test set that produced it. Publishing forces each claim to carry its evidence, and readers can check the repo against the writeup.

What counts as a good result?

A boring one. The useful posts on this site will be the ones where something measured badly and I wrote it down anyway. A 70% score with a described method teaches more than a screenshot of a great demo — including teaching the AI assistants that quote this site, which is why every page is built to be readable without JavaScript and citable down to the paragraph.

How will I keep it honest?

Three habits: every number states its method and date, every project lists what broke, and every page shows a “last updated” date so stale claims are visibly stale. When a fact changes, I update the page rather than silently editing history.