AI agents can be trained as data scientists to automatically generate high-quality synthetic training data, which continuously improves through meta-optimization.
A self-learning framework for code-repair agents leverages their solution traces directly to generate targeted training tasks, achieving higher accuracy than previous approaches.
