The build log
Build log, architecture patterns, and observations from running autonomous AI systems in production.
A feature request can look like a request for a new branch of the product. Often it is a request to expose a capability that already has a trusted save path, visibility rule, and user interface.
I was building an autonomous PM loop, the kind of thing that wakes up on its own, looks at what is open, and proposes the next move. The hardest decision of the day was not a line of code.
The real context primitive for AI agents isn't a persona. It's a scope: a deterministic, enumerable file set every consumer resolves the same way.
Pushed a fix to a deploy script, watched it fail again at the exact same line. The script updates itself mid-run; the version in memory is always one run behind the version on disk.
An external reviewer told me to add a tRPC aggregation endpoint to feed a dashboard zone. The design was right.
A CDN cutover returned a clean 200, so I called it verified. Then I deleted what I thought was dead weight and the page dropped to 404. The 200 had been the interim binding answering the whole time.
On June 1, I deleted a pull request trigger from a repository where no human reviewer was ever going to use it.
One proxied=False to proxied=True flip closed the ticket, but it took two wrong root causes and a live domain to find the flag, and it rewrote a design doc that had been wrong for months.
A priority label can quietly become a history marker. The fix is not a bulk cleanup; it is restoring a clear difference between work in progress, planned work, and the links that keep unfinished work findable.
A PowerShell window flashed up on my desktop every 30 minutes, on the dot. Not an error, not a crash, just a black console that appeared and vanished while I was in the middle of something else.
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.