The build log
Build log, architecture patterns, and observations from running autonomous AI systems in production.
When a workstation slowed under concurrent work, the visible AI workload was an easy explanation. The useful lesson was to measure the whole system, test the hypothesis, and make reversible infrastructure changes that protect data and account for the actual workload.
Chrome was up, CDP answered, automation worked fine, but the window was nowhere. The process was running in Session 0, a Windows desktop no human is ever logged into.
Classic ASP's Request.Form only parses application/x-www-form-urlencoded. Post FormData from a modern fetch call and the collection comes back silently empty with no error. Here is why, and the one-line fix.
An execution-plan operator did not reveal the size of its seek range. In this verified, non-overlapping range dataset, a differently oriented covering index substantially reduced worst-case reads after staged testing.
I was trying to make one surface do too many things. Once I mapped the architecture, the problem was obvious.
An agent harness audit ended with deletions and a refusal to build clone-sync machinery. Most doc-sync tooling is a tax you invented to avoid admitting you have too many copies of the truth.
Off-the-shelf project management tools are built for human teams. My team is mostly AI agents. The tools do not fit.
In one night I fired off 12 parallel Manus tasks. The batch included a 15-question architecture review, 10 parallel system audits, a synthesis pass, and a hosted visual map.
I built 6,134 enrichment sidecars for my multi-agent system and realized absolutely nothing was reading them.
Building a memory layer on a year of agent conversation sidecars: the enricher, the theme aggregator that exposed a 14 percent slice, the nightly schedule, and what shipped unfinished.
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.