We put an agent at the heart of the optimizer. It watches the training dynamics, reacts to them, and rewrites the optimizer’s source code as training goes.