My Dark Factory Reinvented Sprint Planning
After 40B tokens, adding SRE and dreaming loops forced my dark software factory to invent sprint planning, then stalled its TypeScript backend into a Go v2.
~/Code/jduncan-blog
Engineering, craft, and curiosity.
Self-taught engineer, recovering history major. Tech, woodworking, home automation.
After 40B tokens, adding SRE and dreaming loops forced my dark software factory to invent sprint planning, then stalled its TypeScript backend into a Go v2.
Eight days after pruning 65% of Leaf's prompts and adding Dream telemetry: 655 planning rooms cost 40% less, 91% of bugs self-healed, and 1 metric lied.
An accidental 1,199-room randomized trial in Leaf showed Opus 5.5 halves downstream rework as an Architect, GPT-6.1-Sol wins builds, and overrides lose.
At 39.9B tokens and 4,044 runs in Leaf, catching a lying rework denominator proved why 28.3% Tech Lead rejections make 99.6% unattended merges work.
Six agentic development rules from a year of building software factories, 15.8B tokens, and deleting harnesses when the data pushed back.