Optimize agent prompts, code, and trainable parameters end to end using execution traces and general feedback.