Skip to content

Distributed Agent Reasoning Across Independent Systems With Strict Data Locality

🕒 Published (v1): 2025-11-20 12:13 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

A proof-of-concept demonstrating that LLM-driven agents on fully independent organizational nodes (Clinic, Insurer, Specialist) can perform multi-step reasoning and reach coordinated decisions using only natural-language message passing, HMAC-based pseudonymous tokens, and the Orpius OperationRelay mechanism — without any shared schemas, identifiers, or centralized infrastructure.

Problem

Existing multi-agent frameworks assume a single system boundary with shared memory, global context windows, or common tool access. Real inter-organizational workflows (healthcare, finance, regulated industries) require strict data locality: participants cannot share raw records, schemas, or infrastructure. There is no established pattern for autonomous agent coordination under these constraints.

Method

Three independent Orpius platform nodes each host a local LLM agent with access only to node-local CSV files and secrets. Cross-node communication occurs exclusively via OperationRelay, a built-in Orpius mechanism that invokes a remote Operation (a named, permissioned agent endpoint) and normalizes the response into the local tool-calling loop. The information channel is constrained to concise natural-language summaries; the only structured artifact crossing boundaries is an HMAC-derived pseudonymous token:

\[\text{patientToken} = \text{HMAC-SHA256}(\text{key},\ \text{UPPER}(\text{TRIM}(\text{patientId})))\]

The key never leaves the originating node. The data-locality invariant is formalized as disjoint private datasets \(D_i \cap D_j = \emptyset\) and outbound messages satisfying \(D_i \cap m_{i \to j} = \emptyset\), so that each cross-node message is a lossy projection \(m_{i \to j} = f_i(D_i, q)\) and no node can reconstruct another's dataset. The coordination graph is:

\[\text{Clinic} \xrightarrow{m_{C \to I}} \text{Insurer} \to \text{Specialist} \xrightarrow{m_{S \to I}} \text{Insurer} \to \text{Clinic}\]

Key Contributions

  • Architecture for distributed agentic reasoning across fully isolated organizational nodes with no shared schema, memory, or infrastructure.
  • Privacy-preserving cross-organization linkage via deterministic HMAC token (one-way, no identity disclosure, no extra round-trips).
  • Strict data-locality enforcement at the platform level through Orpius node isolation, secret management, and Operation-scoped tool permissions.
  • Federated decision-support model: cooperation emerges from composing local summarization functions \(\{f_i\}\) rather than global state.
  • End-to-end prototype instantiating these patterns in a realistic three-party clinical-coverage workflow.

Results

  • No quantitative evaluation was performed; the paper is explicitly a proof-of-concept.
  • A small number of manual functional test runs verified that the three-node pipeline produces coherent end-to-end outputs (coverage verdict + clinical appropriateness recommendation) on synthetic data.
  • No accuracy, latency, cost, or robustness baselines are reported.

Limitations

  • No clinical validation; guideline extracts and recommendations are synthetic and illustrative only.
  • No formal evaluation: accuracy, latency, cost, robustness, and failure modes are unmeasured.
  • Minimal synthetic dataset; real organizational data is vastly more complex.
  • Natural-language message variability not evaluated across models, temperatures, or prompt variations.
  • Only one coordination pattern (coverage clarification) tested; complex multi-hop workflows unexplored.
  • No adversarial or security testing: prompt injection, malicious summarization, identifier leakage, and cross-system inference attacks are explicitly out of scope.

Relevance to Harnesses / Meta-Harnesses

This work is a concrete instantiation of a distributed multi-agent harness in which the orchestration layer (Orpius) provides node isolation, tool-permission enforcement, and cross-node relay as platform primitives — analogous to what a meta-harness does when coordinating specialized sub-agents. The OperationRelay abstraction is architecturally equivalent to a typed inter-agent dispatch interface in a meta-harness, but stripped of shared schema: each "sub-harness" (node) exposes exactly one entry point and enforces its own tool boundaries. For researchers building harnesses that must compose agents across organizational or trust boundaries — e.g., federating domain-specific sub-harnesses without a shared context window — this paper provides a minimal, formalized blueprint. The HMAC token pattern and the \(m_{i \to j} = f_i(D_i, q)\) message model are directly applicable to designing privacy-respecting inter-harness communication contracts.