I measured why AI coding agents build bureaucracy around their own work, then made a one-file skill to stop it

Disclosure: I built this. Sharing it here as my own work. Across the full git history of one repo an agent built over 20 days, I found the verification machinery (gates, ADRs, traceability docs) grew to 20,280 lines vs 17,964 lines of actual product code, 33%