Original title: Value-Preserving Architectures for Agentic AI Systems
Alessandro Pesare, Tommaso Dolci, Katja Hose, Emanuel Sallinger · September 2026
Systems made of multiple AI agents powered by large language models can automate hard tasks. But they risk violating basic human values like privacy, fairness, and safety. Software engineering usually focuses on making programs run correctly. Today, AI systems also need responsible design that protects these values. The authors argue that how you connect and organize these AI agents shapes how they behave. To fix this problem, they examine three system setups that preserve specific values. First, a privacy-focused setup uses a federated layout where data stays spread out. Second, a distributed network promotes pluralism and diverse views. Third, a guard-agent design monitors the group to spot and reduce unfairness. The team uses real-world examples to show how these designs lay a path toward building trustworthy AI groups.
Imagine working on a group project where every member's notes are shared with everyone else without restriction. Some people might feel their privacy is invaded, and louder students could dominate, ignoring quieter ideas. Here, the researchers propose ways to organize the team's communication—such as keeping personal notes private or assigning a monitor to check that everyone gets a fair say—to prevent these problems in AI systems.
When AI agents are used in real-world systems like healthcare or finance, poor design can lead to privacy breaches or biased decisions. This work shows how to organize these systems to safeguard human values from the start.
The abstract doesn't say whether these architectural patterns have been tested in real-world systems or what challenges they might face in practice. It focuses on proposing the patterns and illustrating them with use cases, but does not address scalability, performance trade-offs, or empirical evidence.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Value-Preserving Architectures for Agentic AI Systems — Alessandro Pesare, Tommaso Dolci, Katja Hose et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.