Let AI suggest before you save
A suggested classification should stay editable until a person confirms it. Separating analysis from persistence turns an AI mistake into a correction opportunity instead of a bad record.
The build log
Notes from running agents in production.
Build log, architecture patterns, and observations from running autonomous AI systems in production.
-
-
A dry run is only useful when it checks the same things as apply
A preview that skips a real precondition can produce a reassuring number and still fail at the moment of change. The solution is to share all decision logic and separate only the final side effect.
-
I Shipped the Kernel, Then Realized the Product Was Wrong
The kernel's v3 reached production with 201 passing tests and a clean architecture. A non-technical user invalidated the interaction model in ten minutes.
-
I Almost Called It a Bot Swarm
The cheap detector said the site was under attack.
-
The bot swarm that turned out to be one bad query plan
The uptime monitor fired on a 9 PM CPU spike. Everything pointed to hostile traffic. It was one cached plan in a request-log table doing clustered scans on every request, and one OPTION (RECOMPILE) settled it.
-
My agents quietly hoarded 48 GB of Chrome profiles until the machine froze
Per-session browser isolation was the right call. Per-session isolation with no reaper filled my system drive, froze the machine, and forced a reboot. The fix was a fixed pool.
-
314 broken stats pages and one bad cursor
Stats pages were broken across 314 sites. The cause was a shared recordset helper opening keyset ADO cursors over aggregate queries. The data caught up.
-
It wasn't one tournament bug
The ticket reopened the same day it closed. A broken admin redirect after new event creation turned into five separate failures across image persistence, custom sports, freeform dates, edit routing, and cart verification, each hiding behind the last.
-
Clearing %TEMP% on a machine running parallel Claude sessions is friendly fire
My C: drive hit 230 MB free out of 231 GB. I cleared the obvious junk and got 18 GB back.
-
The exit survey that overwrote a live mailer
A new exit-feedback page shipped cleanly. The provisioning script had already silently clobbered a scheduled digest mailer row that had been live for months.
Follow the Build
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.