This paper addresses value alignment for agentic AI and large‑language‑model (LLM) based multi‑agent systems (MAS). While traditional software engineering emphasizes functional correctness, MAS architecture—covering coordination mechanisms, communication protocols, and system topologies—directly shapes system behavior and its impact on human‑centered values such as privacy, fairness, and safety.
The authors propose three value‑preserving architectural patterns:
- Privacy‑aware architecture: a federated topology where data stays on‑device and only model updates are shared, reducing centralized leakage risk.
- Distributed pluralism architecture: decentralized nodes and co‑existing diverse models encourage viewpoint diversity and prevent single‑model dominance.
- Guard‑agent architecture: a dedicated guard‑agent monitors other agents, detects and mitigates unfair or harmful outputs, providing real‑time value constraints.
Representative use cases illustrate these patterns: federated learning for cross‑institution medical analytics, multi‑model voting in open‑domain dialogue platforms, and unfairness detection guards in financial risk‑control systems. By embedding human values at the architectural level, the work offers a unified set of patterns and guidelines for building trustworthy MAS.
Review