Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems

The project addresses a critical and under-explored gap in AI safety: emergent bias in swarms. While individual agent alignment (RLHF, DPO) is well-studied, the interaction effects within Multi-Agent Systems (MAS) can lead to 'bias amplification'—a phenomenon where safe individual agents produce prejudiced outputs through collective feedback loops. Technically, this is a research repository (3 days old, 0 stars, 3 forks) and currently lacks the traction or tooling infrastructure to be considered 'defensible.' Its value lies in the methodology and the specific focus on 'topologies' (how agents are connected) as a driver of bias. From a competitive standpoint, frontier labs (OpenAI, Anthropic) are currently more focused on agentic capabilities (OpenAI Swarm, Computer Use) than the specific sociotechnical safety metrics for swarms, giving this research some breathing room. However, once MAS becomes a standard production pattern, platform providers (Microsoft, AWS, Google) are highly likely to integrate similar 'Guardrails for Agents' into their orchestration layers, potentially making standalone measurement frameworks like this obsolete. The risk of displacement is high because as soon as an industry-standard benchmark (like a 'Swarms-HELM') emerges, small research repos are usually absorbed or ignored.