Source-linked AI summary
Value-Preserving Architectures for Agentic AI Systems
Alessandro Pesare, Tommaso Dolci, Katja Hose, Emanuel Sallinger
TL;DR
Agentic AI and MAS enable autonomous, collaborative task execution but make human-centered values harder to preserve as oversight and transparency diminish. This paper links architectural choices to privacy, pluralism, and fairness through three reusable patterns and illustrative scenarios, laying groundwork for a unified catalog of value-preserving MAS architectures.
Problem
MAS can reduce human oversight and transparency, while post-hoc safeguards may allow biased or harmful information to propagate before output inspection.
Method
The paper proposes Federated Silos Coordination, Peer-to-Peer Deliberation, and Plan-Triggered Guard Agents as architectural patterns targeting privacy, pluralism, and fairness.
Results
The patterns structurally support privacy through minimal data sharing, pluralism through distributed deliberation, and fairness through runtime guard agents.
Takeaways & Limitations
Architectural design can serve as a foundation for preserving human-centered values in MAS and developing a unified catalog of trustworthy patterns.
Abstract
from arXiv · showhide
The emergence of agentic AI and LLM-based multi-agent systems (MAS) presents unprecedented opportunities for automating complex tasks, while simultaneously raising critical concerns about the preservation of fundamental human-centered values, such as privacy, fairness, and safety. Although software engineering has traditionally focused on functional correctness, the adoption of LLMs and AI agents into complex socio-technical systems has intensified the need for responsible software engineering and robust value alignment. In MAS, architectural design decisions, such as coordination mechanisms, communication protocols, and system topologies, play a central role in shaping system behavior and the outcomes they produce. This paper argues that architectural choices influence not only the functionality and performance of MAS but can also promote value-oriented system behavior. Therefore, we investigate how different architectural designs support different human-centered values, discussing the following value-preserving architectural patterns: (i) a privacy-aware architecture with a federated topology, (ii) a distributed architecture to promote pluralism and diversity, and (iii) a guard-agent architecture to detect and mitigate unfairness. Finally, we introduce representative use cases to illustrate the proposed architectures in real-world scenarios. By linking architectural design with human-centered values, this work lays the foundation for a unified set of architectural patterns and guidelines towards the design of trustworthy MAS.
1 Introduction
Agentic AI and MAS expand software capabilities through autonomous planning, tool use, and collaboration, but increasing interaction depth reduces oversight and transparency. The paper therefore frames architectural design as a means to preserve privacy, pluralism, and fairness through reusable patterns.
- Agentic AI systems use autonomy, planning, external tools, and runtime feedback, while MAS distribute shared goals across collaborating specialized agents.
- As agent interactions grow, reduced oversight and transparency can allow coordination failures, data-minimization violations, and majority preferences to propagate through the system.
- Post-hoc guardrails may inspect outputs only after biased or harmful information has been introduced, aggregated, and propagated upstream.
- The paper investigates reusable architectural patterns that connect MAS design choices with privacy, fairness, and pluralism requirements.
- It presents Federated Silos Coordination, Peer-to-Peer Deliberation, and Plan-Triggered Guard Agents, respectively targeting privacy, pluralism, and fairness.
- The paper concludes by pointing toward a unified catalog of value-preserving architectural patterns for MAS.
2 Background
The background characterizes agentic AI as autonomous, planning, tool-using software and explains why MAS create behavioral, attributional, and control opacity. Responsible software engineering consequently treats human-centered values as first-class design requirements.
- Agentic AI entities act autonomously, decompose goals into ordered subtasks, and use tools such as APIs, databases, and code interpreters.
- MAS produce behavioral, attributional, and control opacity because emergent interactions, distributed responsibility, and orchestrator-mediated instructions complicate inspection and intervention.
- Responsible software engineering incorporates ethical considerations throughout development and treats privacy, fairness, and explainability as first-class requirements.
- LLM-based software agents have motivated the concept of sociocritical systems, whose ethical and societal consequences require values to be embedded during development.
3 Value-Preserving Architectures
The paper proposes three architectural patterns that structurally preserve different human-centered values in MAS. Federated silos constrain data sharing, peer-to-peer deliberation distributes authority, and guard agents intercept bias during processing.
- The three patterns are specified using context–problem–solution descriptions covering target values, architectural structure, agent responsibilities, and coordination flow.
- Federated Silos Coordination Pattern: Federated Silos Coordination preserves privacy by routing tasks to domain agents that return only minimal task-specific abstractions without exchanging underlying data.
- Peer-to-Peer Deliberation Pattern: Peer-to-Peer Deliberation supports pluralism through direct communication among agents, with collective positions emerging iteratively and no node holding final authority.
- Plan-Triggered Guard Agents Pattern: Plan-Triggered Guard Agents promote fairness by inserting specialized, modular guards into processing flows to detect and correct bias before downstream steps.
4 Patterns in Practice: Application Examples
The paper applies three value-preserving MAS architectures to personalized healthcare, diverse news summarization, and fairness-sensitive hiring workflows. These scenarios show how architectural coordination can preserve privacy, surface minority viewpoints, and trigger human review of biased recommendations.
- 4.1 Federated Silos Coordination for Privacy: Federated coordination limits healthcare agents to clinically relevant information, preventing unrelated legal or financial data from influencing treatment recommendations.The medical agent accesses medical history, medications, and diagnostic results while applying data minimization.
- 4.2 Peer-to-Peer Deliberation for Pluralism: Peer news-retrieval agents independently gather sources and deliberate across differing alignments to represent majority and minority viewpoints.The workflow includes nationalist, green-aware, and minority perspectives before producing the final summary.
- 4.2 Peer-to-Peer Deliberation for Pluralism: Deliberation surfaces alternative perspectives and reduces the risk that centralized summarization over-emphasizes dominant narratives.The architecture distributes retrieval and discussion rather than relying on a single centralized summarization process.
- 4.3 Plan-Triggered Guard Agents for Fairness: Plan-triggered guard agents inspect intermediate hiring rankings when tasks are fairness-critical and identify possible influence from sensitive attributes.The fairness mechanism is selectively activated for bias-sensitive tasks such as resume parsing and candidate evaluation.
- 4.3 Plan-Triggered Guard Agents for Fairness: Human intervention redirects a biased initial ranking toward the more experienced candidate, preventing an unfair hiring outcome.Candidate A is initially ranked above Candidate B despite Candidate B’s better-matching professional experience; the recruiter ultimately selects Candidate B.
5 Conclusions and Research Directions
The paper argues that MAS architectural choices can reinforce human-centered values by design and presents initial patterns supporting privacy, pluralism, and fairness. It identifies composition, evolving requirements, shifting runtime values, and evaluation metrics as directions for future research.
- Conclusion: The paper introduces initial MAS architectural patterns supporting privacy, pluralism, and fairness as a foundation for a unified pattern catalog.The conclusion frames these patterns as a basis for developing human-centered MAS.
- Research Directions: Future work will study how patterns can be composed to satisfy multiple value requirements while examining resulting synergies, trade-offs, and limitations.The planned analysis concerns architectures designed to optimize different values simultaneously.
- Research Directions: Future research will address changing stakeholder requirements and potential shifts in values during system runtime.The conclusion identifies runtime adaptation to evolving requirements and value changes as an open problem.
- Research Directions: Defining metrics for evaluating value preservation remains a crucial challenge for operationalizing human-centered values in software systems.The paper gives pluralistic alignment metrics as an example of this evaluation need.