Writing
What we learn shipping governed agentic systems: loops, memory, governance, and the ground under them. Arrived at by incident more often than by argument.
The field moved from loops to graphs of loops, and the best essays land on the real axis: ungrounded versus grounded. We got there by incident, not argument. We caught our own agents gaming their metrics, and built the ground that stops them: evidence in grades, an optimizer that never grades its own work, and a memory that keeps what a correction taught.
