Value-Preserving Architectures for Agentic AI Systems | AI Research

Value-Preserving Architectures for Agentic AI Systems
As AI agents and multi-agent systems (MAS) become more common in software development, they often operate with limited human oversight. This creates a significant challenge: how to ensure these systems remain fair, private, and pluralistic. This paper argues that we cannot rely solely on "post-hoc" fixes—such as filtering outputs after they are generated—because biased or harmful information often propagates through the system long before it reaches the final output. Instead, the authors propose that human-centered values must be built directly into the system's architecture.

Rethinking Architectural Design

The authors identify that the way agents are organized—their communication protocols, topologies, and coordination mechanisms—directly shapes the system's ethical outcomes. When agents are left to interact without structural constraints, they suffer from "opacity," where it becomes difficult to trace why a specific decision was made or where a bias originated. By treating values like privacy and fairness as core architectural requirements rather than afterthoughts, engineers can create systems that are inherently more trustworthy. The ai agents story also surfaces in Stanford Researchers Develop TRACE to Fix..., adding another angle.

Three Patterns for Value Preservation

The paper introduces three specific architectural patterns designed to embed human-centered values into the structure of a multi-agent system:

• Federated Silos Coordination (Privacy): This pattern uses a hub-and-spoke model where a central coordinator manages tasks across different domains. Crucially, domain-specific agents never share data with each other. They only send minimal, abstracted information to the coordinator, ensuring that sensitive personal data remains contained within its original domain.

• Peer-to-Peer Deliberation (Pluralism): To avoid the bias of a central authority, this pattern removes the orchestrator entirely. Agents with distinct perspectives communicate directly with one another to reach a consensus. If they cannot agree, the system surfaces the disagreement to the human user, ensuring that diverse viewpoints are represented rather than filtered out.

• Plan-Triggered Guard Agents (Fairness): This approach inserts specialized "guard" agents into the workflow. Before a system executes a task, these guards review the proposed plan for specific biases—such as those related to gender or ethnicity. If a bias is detected, the guard intervenes before any action is taken, ensuring fairness is maintained throughout the process. The ai agents story also surfaces in Andrew Ng Launches OpenWorker to Deliver..., adding another angle.

Moving Toward Responsible Engineering

The proposed architectures represent a shift toward "responsible software engineering." By separating the system into distinct phases—such as planning, review, and execution—the authors provide a framework that allows developers to update or replace specific components without rebuilding the entire system. This modularity makes it easier to maintain ethical standards as the system evolves.

Practical Application

To demonstrate these concepts, the authors apply their patterns to real-world scenarios, such as medical decision-support systems and news aggregation. In these examples, the architecture acts as a structural guarantee: in the medical scenario, the Federated Silos pattern prevents financial data from influencing health recommendations, while in the news scenario, the Peer-to-Peer pattern ensures that multiple viewpoints are considered. These use cases illustrate that by linking design choices to specific human values, engineers can retain better control over the behavior of complex, autonomous AI systems. The ai agents story also surfaces in Alibaba Releases Page Agent to Control..., adding another angle. as detailed in the full paper on Arxiv