There is a large potential for AI to streamline complex decisions, like coordinating between a doctor’s clinic and an insurance provider. But there’s a massive data wal in the way: Privacy.
We recently published a paper, "Distributed Agent Reasoning Across Independent Systems With Strict Data Locality," which demonstrates a way to break through those walls. We’ve built a system where AI agents from different organizations can collaborate to solve a problem without ever seeing each other’s private data.
The Problem: Data That Cannot Be Shared
In a perfect world, if you want an AI to help with a medical insurance claim, you'd simply feed all the data—patient records, insurance policies, and clinical guidelines—into one big pot where a model can reason over it.
However, regulatory walls make this impossible. Patient data cannot simply be moved, merged, or exported.
Eveven though multiple organizations possess the information needed to solve a problem, no single system can legally or operationally access all of it. This creates a deadlock where high-value automation stalls because the data is stuck behind a firewall.
The Orpius Solution: Reasoning Instead of Sharing
Instead of attempting to centralize data, Orpius allows for a distributed network of specialized agents. In our research, we modeled a three-way interaction between a Clinic Agent, an Insurer Agent, and a Specialist Agent. The breakthrough isn't just that they talk to each other, it’s how they collaborate:
Strict Data Locality: Every agent lives on its own "island." No raw records ever leave their home server.
Digital Handshake: To ensure the agents are talking about the same case without knowing who that person is, we use a unique "digital handshake." They can verify they are looking at the same file without ever exchanging a name or ID.
Natural Language Summaries: Instead of exchanging rigid, data-heavy tables, our agents communicate via concise natural language. This need-to-know messaging prevents accidental data leakage while keeping the reasoning human-readable.
Does it actually work?
We showed that these agents could successfully navigate a Clinical-to-Insurer workflow.
The Clinic identified the medical need.
The Insurer verified the anonymous patient's plan.
The Specialist gave a thumbs-up on the clinical appropriateness.
The Final Verdict was delivered back to the user all without a single piece of PII (Personally Identifiable Information) crossing a boundary.
Why This Matters
This isn't just a healthcare fix. This architecture provides a blueprint for any industry where data is siloed. Imagine:
Supply Chains: Multiple companies coordinating logistics without revealing their full inventory to competitors.
Banking: Detecting fraud across different institutions without sharing private customer lists.
We are moving away from the era of "Centralized AI" and toward Federated Reasoning. It’s a world where AI doesn't need to "own" your data to be helpful, it just needs to know how to ask the right questions.