Notes from running AI coding agents on production work — mostly about the gap between a practice being written down and a practice actually happening.
What still reproduces on a 4B local model at pinned versions — and one result of mine that did not survive re-running, which turned out to be the useful part. With scripts you can run.
I measured whether my agents followed my model-routing rule. They didn’t — 0.03% of output. So I moved the rule below the decision. A month later I found the counter itself had been blind to the one surface that mattered.
Two independent agent sessions needed to message each other. Relaying copy-paste myself failed in two specific ways, and neither was a typo — which turned out to be the argument for a channel with provenance.