Auditable autonomy is becoming a practical requirement for deploying large language model (LLM) agents in operational workflows where recommendations can trigger consequential actions. Many autonomy claims remain hard to evaluate because studies emphasize task completion or fluent explanations while underreporting tool privileges, verification conditions, rollback feasibility, and trace completeness. This review develops a decision-making–centered framework that treats autonomy as an auditable engineering property. It introduces a three-plane big data foundation: an evidence plane with provenance and freshness constraints; a decision-trace plane that records retrieval identifiers, tool invocations, intermediate checks, and policy evaluations; and an outcomes plane that captures post-decision effects such as KPI shifts, rollback incidence, recurrence, and operator overrides. On this basis, we present an assurance stack taxonomy covering grounding controls, verification and precondition checks, action gating for risk bounding, and accountability mechanisms. We then propose an evaluation ladder that moves beyond text metrics toward evidence fidelity, action validity, and longitudinal stability, with reporting requirements that improve cross-domain comparability. Finally, we synthesize accidental failures and adversarial threats, including staleness, drift, prompt injection, partial execution, and trace tampering, and map trace-based detection signals and mitigations to the assurance stack.
Theodorakopoulos et al. (Thu,) studied this question.