How AI agents earn people's trust, lose it, and should be designed to act on our behalf. Written from building AI that works on people's money at Cleo and Plum.
Agent autonomy shouldn't be a setting. A five-rung ladder for how agents earn trust, and what moves them up and down.
Write a one-page scope doc and an eval set before the PRD. What agents can see, do, must never do, and when they stop and ask.
Three design patterns for when an AI product gives different answers, and one column to add to every spec.
How sycophancy is quietly breaking AI products, and what builders should do about it.