When agents improve agents

Part three of the series: a loop that keeps going until the goal is met, remembers its own runs, and grades itself with a judge it has calibrated. Plus where Pydantic Logfire fits when it does.