The build log
Build log, architecture patterns, and observations from running autonomous AI systems in production.
A SQL Server execution plan showed Index Seek. The same query was doing 8,000 logical reads. Here is what the operator name hides, and how a second covering index collapsed the worst case from 14,718 reads to 3.
I sent a cross-cutting feature design to a read-only AI reviewer three times. Each pass found new, real bugs: eleven findings total, none hallucinated. Here's what made the difference.
Thread traceparent headers from Classic ASP into a modern Node collector via IIS header forwarding, so a sidecar emits spans on behalf of the legacy layer without touching its core logic.
The query plan said Index Seek. The query was doing ~8,000 logical reads per call against a 4-million-row table and burning ~330ms of CPU under load.
I needed a Cloudflare tunnel to survive reboots without me babysitting it, so I wired it up to autostart and walked away.
I verified the SQL Server edition before shipping an ONLINE index build. Still shipped the wrong answer. The only thing that caught me was the server refusing the statement.
Every push to my internal API repo showed a red failure, so I assumed CI was running and just unhappy with my code. That assumption was wrong twice over.
On May 17, a security review put one humiliating string at the center of our portal: ?k=<token>.
A single environment variable was supposed to finish a drive migration. The folder rename still failed with Access Denied. The real blockers were background services I had forgotten were running.
A CAD customer paid a sum. Staff saw a sum no label. PayPal had not lost anything. Our capture handler stored the USD settlement and discarded the buyer-currency breakdown.
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.