MediRush-SafeAgent: Layered Runtime Enforcement for Healthcare-Commerce LLM Tool Execution
Sham Satish Thakare
Abstract
Autonomous large language model (LLM) agents executing multistep tool calls in healthcare-commerce environments present catastrophic failure risks when actions cannot be reversed. In this paper, we introduce MediRush-SafeAgent, a layered runtime boundary enforcement framework designed to guarantee safety invariants during high-stakes agent execution. We establish deterministic state interception, schema-aware pre-execution guards, and out-of-band validation barriers that decouple agent decision proposals from irreversible system mutations. Empirical evaluations across clinical ordering and fulfillment scenarios demonstrate zero invariant violations with minimal latency overhead.
Artifacts & Verified Links
BibTeX
@article{thakare2026medirush,
author = {Thakare, Sham Satish},
title = {MediRush-SafeAgent: Layered Runtime Enforcement for Healthcare-Commerce LLM Tool Execution},
journal = {Intelligence-Based Medicine},
year = {2026},
note = {Under Peer Review, Manuscript ID: IBMED-D-26-00573; SSRN Abstract: 7446543},
url = {http://ssrn.com/abstract=7446543}
}