Agentic AI-Assisted Governance: Towards Semi-Autonomous State Oversight

Abstract

The detection of irregularities in public administration often takes weeks or months when it occurs at all. This article proposes a multi-agent framework based on agentic Artificial Intelligence (AI) for continuous, semi-autonomous, and systemic government oversight. The framework was developed from a literature review on agentic AI, algorithmic governance, and state oversight, and was structured into four functional layers and five specialized agents coordinated by a differentiated escalation mechanism that preserves human decision-making for the highest-impact findings. It was validated through a workshop involving 12 experts in public administration, AI, and auditing, applying Lawshe’s method and the MaxDiff technique to simulate scenarios of contract splitting and concealed budget diversion. The framework detected irregularities within 6 to 9 minutes, compared with weeks required by conventional audits, and achieved a content validity index of 0.85, supporting the relevance and consistency of the proposed framework. Thus, the study represents one of the first multi-agent reference frameworks to explicitly integrate agentic AI capabilities for continuous and semi-autonomous state oversight, applicable to the Latin American context, and opens future research avenues on pilot implementations, the legal framework governing evidentiary admissibility, and oversight of the auditing system itself.