A Stanford framework that treats prompting as an optimization problem, allowing agents to self-improve their instructions over time. Declarative, self-improving language model pipelines.