A javax/Jetty application, a payments-style compliance clock, and a build log that includes the claims we got wrong before we got them right.
We have derived the same capability-routing architecture three times, at three different tiers, in three different languages. The selection logic was never the hard part. What decides whether a distributed fleet exists at all is a question most architectures answer by accident.
Ten rule files govern our AI agents. Not one was written from first principles — each section exists because something went wrong on a specific date, and the date is in the file.
Not because it was expensive, and not because the agents were bad. Because a library gets copied, and a copy cannot be specific — our dungeon-crawler project is still carrying a Windows Hello integration specialist.
Two unsolicited recruiter emails arrived the same week. Neither turned out to be a scam. Working out why they arrived was far more useful than working out whether to trust them.
One five-line SQL function decides whether forgetting to set a variable leaks every tenant's data or shows you none of it. We know which, because ours fired by accident and took down our own sitemaps.
The most reliable guardrail in our agent fleet is not a prompt. It is the list of tools each agent was never given — and the two agents we gave everything to are the interesting part.
Three providers in two years: one grant aged out, one lapsed and the project was deleted outright, and the current one expires on a date we have written down. The balance was never the number that mattered.
Four hundred CI runs failed with no logs, no jobs and no error message. Two experiments, four lines each, proved where the problem was not — which is the only thing that stopped us editing YAML for a week.
Our monitoring called a site healthy for eight days while it served a broken page. Fixing that is the easy half. The hard half is that a health check which only tests success cannot see an authorization failure at all.
Documentation in a fast-moving repo does not go stale gracefully. It goes stale silently and then gets cited. Our fix is a file that ends by naming which other files it beats — and the day it failed is the useful part.
We had to prove a dozen cloud resources were deleted with no console, no billing data and no credentials. A DNS query did it — but only in one direction, and the other direction is where the vulnerability lives.