AI-Agent Governance Must Shift from Model Outputs to Behavior
A Brookings analysis adapted by Internet Law Review argues that governance must move from reviewing model outputs to overseeing the actions of autonomous AI agents that plan, call tools and access external systems. It recommends structured authorization, least-privilege access, end-to-end logging, human confirmation for high-risk operations and runtime monitoring. The authors caution that no single control is sufficient and that existing risk-based, lifecycle governance must become more precise as agents act faster with less human intervention.
As agents gain permissions to send email, modify records, place orders, transfer funds and call APIs, the report argues that authorization, audit trails and runtime controls become central to AI governance.
过去两年，人工智能的前沿发展，正在从“会回答的模型”变成“会办事的智能体”。这里的智能体（AI agents），指能够自主规划、调用工具、连接外部系统并执行任务的技术系统。2025年，智能体在计算机任务上的准确率从12%提升到66.3%，距人类平均水平只差6个百分点；市场研究机构Precedence Rese......
本文来自微信公众号： Internet Law Review ，作者：Landry Signé
过去两年，人工智能的前沿发展，正在从“会回答的模型”变成“会办事的智能体”。
这里的智能体（AI agents），指能够自主规划、调用工具、连接外部系统并执行任务的技术系统。2025年，智能体在计算机任务上的准确率从12%提升到66.3%，距人类平均水平只差6个百分点；市场研究机构Precedence Research预计，全球智能体市场将从2025年的80亿美元增长到2034年的1990亿美元。
#Brookings Institution#World Economic Forum#Precedence Research#NIST AI RMF