Beyond Single AI: The Rise of Multi-Agent Orchestration A Survey on Bias, Privacy, Robustness, and Interpretability

Large Language Models (LLMs) are increasingly deployed within multi-agent systems to enable structured, multi-turn coordination among interacting AI agents. While such orchestration improves reasoning capacity and task specialization, it introduces emergent reliability risks that extend beyond single-agent evaluation, including bias amplification, adversarial vulnerability, data leakage, and decision opacity. This survey systematically analyzes multi-agent LLM orchestration through four interdependent reliability dimensions: bias, privacy, robustness, and interpretability. We introduce an operational taxonomy that categorizes existing systems into debate-based, workflow-based, and hybrid paradigms, and examine how coordination structure shapes reliability trade-offs across these dimensions. Beyond summarizing recent advances, we identify structural fragmentation in current audit tools, benchmarks, and orchestration frameworks, showing that no unified evaluation ecosystem presently captures emergent, cross-agent risks in multi-turn settings. We conclude by outlining research directions toward integrated, automated evaluation environments for secure, ethical, and interpretable multi-agent AI systems.