Agentic AI Trustworthiness Is a Full-Stack Problem: A Framework from Hardware to Governance

Abstract

Multi-Agent Systems (MAS) enable autonomous agents to reason, collaborate, and interact with complex environments. Their increasing autonomy and interdependence introduce system-level trustworthiness challenges that cannot be addressed by securing individual agents alone, as failures originating at lower system layers may propagate across agent boundaries and affect system-level behavior. This paper proposes a Full-Stack Trustworthiness Framework that organizes MAS trustworthiness concerns into six interconnected layers: infrastructure, model optimization, model, cognition, execution, and socio-technical interaction. The framework maps layer-specific threats to corresponding defenses and emphasizes both cross-layer and cross-agent failure propagation. It is evaluated through fault injection into RRAM-based Computing-in-Memory (CiM) hardware executing a three-agent MuSiQue question-answering pipeline built on a Large Language Model (LLM) backbone. The evaluation considers stuck-at faults (SAF) and temperature-dependent weight drift injected into an upstream agent, while downstream agents operate on clean hardware. The results demonstrate that hardware-level faults can propagate across system layers and agent boundaries, resulting in substantial degradation of downstream agent performance. Fault-aware training using a Sparse Mixture-of-Faults (Sparse-MoF) adapter bank improves resilience under both fault conditions. These findings demonstrate the need for full-stack assessment and cross-layer defenses in the development of trustworthy MAS.