Our paper introduces AgentBreeder, a framework for multi-objective self-improving evolutionary search over scaffolds, specifically targeting scaffolds' safety impact on large language models in multi-agent systems.