John Feminella recounts the 2020 Citibank payment error in which loan principal went to lenders when only interest was intended. He argues that warnings and human approval can fail when the underlying action is misunderstood.
Feminella describes an agent registry that records ownership, purpose, tools and endpoints, together with risk-based review that distinguishes low-risk experiments from agents that can affect production.
The talk proposes enforcing tool and resource accessAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use. according to an agent’s declared purpose, and providing shared evaluation infrastructureEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. so teams can test changes without rebuilding common tooling.
Finally, Feminella recommends using conservative runtime classifiers for routine actions while escalating ambiguous or dangerous decisionsHuman judgment in AI is the accountable interpretation and decision-making people contribute when setting goals, evaluating evidence, and managing consequences.. He presents the controls as complementary rather than a substitute for judgment.
Watch on YouTube
