Separating Evidence from Executability: A GraphRAG and Deterministic Validation Architecture for Autonomous Agents

Abstract

Autonomous LLM agents are increasingly issued tool calls that mutate real systems provisioning infrastructure, moving funds, updating clinical records yet most agent architectures conflate the retrieval that informs a decision with the executability of the action that follows, and the resulting probabilistic substrate leaks even when the model's own reasoning correctly identified the risk. This paper proposes an architectural separation. An evidence plane, implemented as a GraphRAG store, supplies the planner with typed entities and relationships in the ordinary way; a deterministic validation plane re-queries the same graph and decides, per proposed tool call, whether it is executable. Probabilistic reasoning proposes an action, while deterministic software decides whether that action is executable. We formalize the interface as an executability predicate composed of schema, permission, precondition, and graph-fact consistency subpredicates, and we specify a mapping from check outcomes to four action paths accept, reject, correct, and escalate. The contribution is a design-science artifact: a reference architecture, a formal interface, and a JSON/YAML tool-and-rule specification deployable behind an MCP-compatible registry, illustrated through worked synthetic examples covering each of the four action paths.