Integrating Electronic Health Records (EHRs) into clinical workflows is crucial for advancing healthcare delivery but poses significant challenges, especially in improving humanmachine interactions through natural language queries. This study builds on prior research [1] by introducing a multi-agent system that uses Large Language Models (LLMs) for secure interactions with healthcare data. In this extended work, we present a new capability for real-time updates to patient records through the addition of a Data Update Agent (DUA), ensuring privacy, accuracy, and compliance with regulatory standards. Compared to prior work, the system features a dual-pathway design for distinguishing between data retrieval and updates, enhanced modularity for seamless agent upgrades, and robust mechanisms to manage complex scenarios and noisy inputs. These advancements improve scalability, fault tolerance, and adaptability to real-world clinical environments. Comprehensive evaluations have been conducted using diverse clinical scenarios, including tests with noisy inputs and complex queries. The results highlight the system’s scalability, accuracy, and practicality, demonstrating its superiority over baseline methods. The proposed framework enables better integration of LLMs in clinical settings by bridging natural language interfaces with secure, interoperable healthcare data systems.

Modern data centers underpin smart-city services, yet their control loops must reconcile noisy, missing, or deliberately deceptive telemetry before acting autonomously. We present a two-path agentic architecture in which a deterministic fast path fuses sensor reports through a reliability-based fuzzy-preference OWA operator (FPR–OWA) and a persistent consensus-reaching process (CRP), while an optional reactive or LLM supervisor can only propose typed actions that a deterministic verifier must admit. Across 1890 indexed simulation scenarios and seven attack families, no estimator dominates: averaging attains the lowest global error and sensor redundancy explains most of the accuracy variation, whereas FPR–OWA + CRP provides the strongest anomaly-diagnostic signal. We prove that logistic dominance preserves the reliability ordering and add a zone-dynamic temporal-consistency ablation. In the original safety campaign, no unsafe execution was observed in the verifier-gated episodes, and the unified process-local runtime later passed all 24 deterministic fault-injection cases; both are empirical, not formal, guarantees. An illustrative closed-loop pilot adds standard control metrics and reveals delayed recovery under common-mode sensor bias; a CPU benchmark keeps persistent-CRP median latency below $0.30$ ms at 100 sensors. The result is a reproducible, bounded blueprint for verified autonomy in datacenter infrastructure; hardware validation, production authentication, and energy measurement remain future work.

Artificial intelligence (AI) agents need authenticated identities, bounded action scopes and evidence before receiving adaptive authority. We present a post-quantum trust, accountability and control (TAC) protocol composing signed requests, independently signed observations, scoped admission and transactional evidence consumption. A Tamarin abstraction checks six universal receipt properties and two existential witnesses; three mutants expose omitted mechanisms. Replay preserves 38,400 recorded controller positions. Execution study E6 covers 48 matched configurations and eight ablations: the bound dispatcher produces no unpermitted effect in 22 mediated configurations, while both privileged bypasses succeed. Twelve-visit study E7 compares adaptive authority with competent static enforcement in 28 matched episodes, alongside 24 isolated execution ablations and two history-mutant diagnostics. Both policies resist tested command substitutions; adaptive warm-up withholds 24 of 48 compliant proposals. False authorized reports violate the illustrative task predicate under both. Two diagnostics show signed-prefix reconstruction detecting injected duplicate reinforcement before effects, and a separate arithmetic oracle checks retained trajectories. Paired-path medians are 3.41 ms for the module-lattice-based digital signature algorithm ML-DSA-44 and 2.19 ms for Ed25519 on the recorded x86 configuration; separate primitive experiments examine build sensitivity. Deterministic results establish conditional composition boundaries with progress costs. Truthful observation, complete mediation, provisioning and durable storage remain assumptions.