Argues that sequences of modest agent tool calls can produce consequential outcomes and shift informational power away from institutions.
The article argues that while individual tool calls may appear modest, their sequential execution can produce consequential administrative outcomes (such as suspending benefits or reallocating resources) that exceed the practical meaning of narrow software permissions. This agentic intermediation shifts informational power away from underlying institutions because users and officials interact with an autonomous intermediary rather than the institution itself, creating new risks of authorization drift, attribution fragmentation, and accountability gaps.
The article treats agentic AI as a governance and accountability problem for the administrative state, not merely a technical deployment issue. Its central argument is that AI agents can produce significant institutional effects through sequences of individually modest tool calls—such as querying records, drafting communications, updating case files, invoking APIs, or triggering workflows—where the cumulative outcome may be consequential even if no single action appears high-risk. In public-sector contexts, this matters because administrative power depends on traceable authority, procedural fairness, and institutional knowledge; when autonomous or semi-autonomous systems mediate those processes, the state may lose the ability to reconstruct what happened, why it happened, and who is responsible.
A key insight is that agentic systems can shift informational power away from institutions. The article emphasizes that oversight becomes harder when actions are distributed across models, tools, vendors, logs, and human approvals, creating opacity about intent, provenance, and causal responsibility. It therefore links cybersecurity accountability—such as monitoring, auditability, incident response, and least-privilege access—to broader administrative accountability, including due process, transparency, and public trust. Human oversight is framed not as a formal checkbox but as a substantive control: agencies must be able to intervene, verify, explain, and reverse agent-driven outcomes where appropriate.
The material matters because administrative agencies are increasingly using AI to support high-stakes functions, from eligibility determinations and case management to security operations and service delivery. Without governance structures that anticipate compounding agent behavior, institutions risk embedding opacity and speed into public decision-making while weakening their ability to assign responsibility. The article’s contribution is to make explicit that the risk of agentic AI in government is not only about model error or malicious misuse, but about the erosion of institutional knowledge, accountability, and control over the informational foundations of administrative power.