Writing
How autonomous systems fail quietly
Everything here comes out of a loop that has been running over my own projects for more than 250 cycles, with an append-only log of what it got right and what it got wrong. No hypotheticals — every incident has a date and a measurement.
Your AI agent's tests are lying to you
I corrupted a data table on purpose to prove my checker worked. It reported clean. Five distinct ways a test can run, pass, and verify nothing — and the ten-second experiment that finds all five.
Build an AI system that can be proven wrong
A system that records only its successes cannot discover it has started failing. The fix is one field and one counter-intuitive rule: publish the hit rate, never gate on it. Mine has fallen from 71% into the fifties — and the piece now carries an update on what happened when it started climbing again, which was worse news than the fall.
If your team is running agents
I build and audit autonomous AI systems. If you cannot tell verified from skipped in your own pipeline, that is the conversation I would like to have.