Blog
Notes on building a commitment device and governing an AI coding fleet.
RSSFive AI models took our code review exam. The newest one would not stop talking.
We replayed seven real bugs from our own history against five coding models and graded them blind. The GPT models won the review, the Claude models found bugs and shipped them anyway, and the newest Opus blew through a word limit half the time.
September 22, 2026 · 4 min readEngineeringThe night we moved CI off GitHub-hosted runners
GitHub put a billing hold on our Actions runners with nothing owed. We took the hint: one line per job, one night, and a gate that finishes in half the time on runners the hold could not reach, for a projected 85 percent less.
September 16, 2026 · 3 min readEngineeringThe Assistant Could Not Have Known
Our in-app AI confidently told us a website was blocked. It was not. The fix was not a smarter model: it was two lost facts and one property we would defend hardest.
August 20, 2026 · 4 min readEngineeringEnforce, don't instruct: governing an AI coding fleet
We build Ulyssify with a fleet of AI coding agents. Here is how we give them judgment about which decisions are theirs, and why we put the ones that matter behind mechanisms instead of words.
August 15, 2026 · 3 min read