The article argues that effective delegation for AI agents depends on a capability-based authorization model rather than traditional identity and access management (IAM) systems. The author begins by highlighting the inadequacies of IAM, which often relies on static permissions tied to identity. This approach fails to account for the dynamic nature of tasks that agents engage in, where authority needs to adjust based on current intent and context. The author uses the analogy of a valet key versus a master key, emphasizing that most systems operate like handing over a master key, which doesn't limit access or enforce specific constraints. The piece identifies three primary issues with current IAM systems. First, while IAM checks attributes like identity and resource tags, it doesn't verify the delegation of authority itself. Second, IAM binds permissions to identities, ignoring that agents operate in real-time and may need to adapt their authority mid-task. Finally, the "Confused Deputy" problem arises when agents misuse authority because they lack context about its origin. These failures stem from IAM’s focus on identity rather than authority, which cannot adequately capture the nuances of delegated tasks. The author proposes a capability token model as a solution. Each token explicitly defines the authority passed, including the actions allowed, constraints, and a history of delegation, ensuring that authority only diminishes as it’s handed down. For example, a capability token might specify that an agent can read specific CSV files, with defined expiration and issuer chains. This structure prevents overreach and maintains clear provenance of authority, addressing issues like the Confused Deputy by ensuring agents only act within the limits of what has been explicitly delegated to them.