Fairness in Multi-Agent Systems for Software Engineering: An SDLC-Oriented Rapid Review

AI systems in software development lack fairness oversight: The review found that transformer-based LLMs and multi-agent systems are increasingly used throughout the software development lifecycle (from code writing to review to release), yet their fairness implications remain largely unexplored despite their significant influence on what code gets created and deployed.

Fairness harms are concrete and varied: Reported problems include representational harms (certain groups underrepresented), quality-of-service failures (unequal tool performance), security and privacy breaches, and governance failures—with some harms concentrated in specific stages of the development process where evidence is weakest.

Current evaluation methods are fragmented and incomparable: Across the 18 studies analyzed, fairness is measured inconsistently using different metrics (accuracy on bias benchmarks, demographic disparity measures, and emergent concepts like "bias amplification"), making it difficult to compare results or determine which approaches actually work.

Real-world deployment readiness is questionable: The review identifies three critical gaps: evaluation practices aren't tailored to multi-agent systems specifically, studies use simplified environments that don't reflect actual software workflows, and proposed solutions for fixing fairness problems are rarely tested in real deployment contexts.

Governance mechanisms for fair AI in software development are underdeveloped: There is scarce evidence for practical mitigation strategies that align with how software teams actually work, suggesting the field cannot yet reliably guide organizations deploying these systems responsibly.