AgentBreeder: Mitigating the AI Safety Risks of Multi-Agent Scaffolds via Self-Improvement

TL;DR

AgentBreeder is a framework that uses evolutionary search to improve multi-agent scaffolds, significantly enhancing safety performance while balancing capabilities, and highlighting potential risks of multi-agent systems.

Contribution

It introduces a novel self-improving evolutionary framework for multi-agent scaffolds, addressing safety concerns and demonstrating effectiveness on benchmark tasks.

Findings

79.4% safety performance uplift in 'blue' mode

Emergence of adversarial scaffolds in 'red' mode

Framework balances safety and capability improvements

Abstract

Scaffolding Large Language Models (LLMs) into multi-agent systems often improves performance on complex tasks, but the safety impact of such scaffolds has not been thoroughly explored. We introduce AgentBreeder, a framework for multi-objective self-improving evolutionary search over scaffolds. We evaluate discovered scaffolds on widely recognized reasoning, mathematics, and safety benchmarks and compare them with popular baselines. In "blue" mode, we see a 79.4% average uplift in safety benchmark performance while maintaining or improving capability scores. In "red" mode, we find adversarially weak scaffolds emerging concurrently with capability optimization. Our work demonstrates the risks of multi-agent scaffolding and provides a framework for mitigating them. Code is available at https://github.com/jrosseruk/AgentBreeder.