AI agents need identity, least privilege and human approval to work safely

The stages include issuing a workload identity, authenticating every tool call, granting task-scoped privileges, requiring approval for consequential actions, and recording the principal and result. The article warns that shared API keys—where every agent has equal standing—misrepresent AI agent identity by removing critical boundaries. It also highlights risks like silent expansion of authority as tools and credentials accumulate, stressing the need for controls that act before irreversible consequences occur. Evaluation should focus on measurable outcomes, failure modes, and recovery mechanisms rather than polished demonstrations.