Writing

Notes from the engineering work. What breaks when AI-generated systems meet production, how to make agent collaboration hold across sessions, and the infrastructure it takes to make agent output reliable.

• Telling an agent to be honest isn't enoughA memory rule wasn't enough to stop Claude Opus 5 from turning quietly agreeable in deep sessions. On the four-leg doctrine I put in place instead, and a July paper that reframes what I built. August 2026.