Title: Ensuring Safety and Auditability in LLM Agents Through Human-Created Ontologies

Abstract

This paper presents Agentic Redux, a novel LLM agent architecture designed for complex problem spaces that demand linear auditability. By leveraging the typed lambda calculus, we demonstrate that when deployed in suitable domains, Agentic Redux ensures semantic correctness in its executions. Furthermore, every decision made by the system is permanently logged in an append-only ledger. We highlight two production-ready applications of this architecture: healthcare billing compliance and security vulnerability disclosure. Complete, functional code for Agentic Redux operating within both these contexts is provided in an accompanying repository. Additionally, we propose Ontology-First Agent Design, a framework for building agent systems tailored to specific problem areas. This methodology involves a human expert mapping the domain using Basic Formal Ontology, followed by the assignment of an LLM to determine the specific roles that both agents and human-in-the-loop participants can assume to address challenges within that domain.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC