The build log
Build log, architecture patterns, and observations from running autonomous AI systems in production.
At 4:06 PM on May 8, I merged the checkout feature branch into the Packages redesign branch, pushed to an internal host, and sent the reviewer a URL. That part was routine. What was different was the page he loaded.
The pings were clean. Zero packet loss to anywhere I tested, DNS resolved in milliseconds, and a TCP handshake to port 443 completed fast.
Intermittent Playwright failures across parallel Claude Code clones traced to one shared Chromium profile lock. The --isolated flag removed the contention.
claude --resume <sid> worked when I ran it in my terminal and returned No conversation found when my web app ran the exact same command, on the same machine, against a session file that demonstrabl...
Building a conversation archive across four provider repos is the easy part. The useful part was sending the spec to Manus and watching it find gaps in every assumption.
Elapsed time, concurrent process time, human attention, and verified outcomes are different measures. Once work happens in parallel, a useful record must preserve provenance, uncertainty, review state, and the limits of what any productivity metric can claim.
I asked an agent to set up a scheduled job that would keep building out my agent-log archive.
AI made ticket analysis nearly free, which exposed the real cost: a read-only backlog page. Inline verbs turned tab-thrash into one-click decisions.
A morning IPv6 tweak made localhost resolve to ::1 and every request hang for ten seconds. Timed curl probes found the dead hop in a three-hop proxy chain.
Rebuilding a work-tracking dashboard around Area, Epic, and Task instead of sessions. What WAIT_HUMAN detection unlocked, and the tax of renaming 19 directories to strip dates from IDs.
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.