Provably Auditable and Safe LLM Agents from Human-Authored Ontologies

Aaron Sterling

We introduce the LLM agent architecture Agentic Redux, intended for use with nontrivial problem domains that require linear auditability. Using the typed lambda calculus, we prove…