AgentBreeder: Mitigating the AI Safety Risks of Multi-Agent Scaffolds via Self-Improvement | Read Paper on Bytez