BUILD AN AGENT LOOP A SMALL MODEL CAN FINISH

Summary from the TLDR Dev newsletter

A reliable agent loop depends less on model size than on a trustworthy check that turns progress into a score or pass condition. In two visual tasks, small models improved dramatically when they received repeatable feedback across varied examples and holdout cases, while weak checks encouraged overfitting or shortcuts.