Agent Approval Fatigue

Agent approval fatigue is the source’s practical name for what happens when users or teams have to approve too many small agent actions. In 「模型能力已经够了，要卷就卷 infra」｜对谈戴冠兰：Runta 创始人, Koji describes becoming more willing to let an agent access Gmail as trust increases, while [[DaiGuanlan|戴冠兰]] argues that permissions should often be granted only for a specific task and revoked after completion.

The concept extends Agent Permission Boundaries by adding the human-behavior failure mode. If every email read, credential access, file edit, or API call requires approval, the agent becomes too slow to be useful. If the user responds by granting broad standing authority, the blast radius expands. A good Agent Runtime Execution Layer therefore needs temporary scopes, audit trails, risk-tiered actions, and escalation rules that reduce interruptions without turning convenience into unmanaged authority.

Key Claims

• Approval fatigue is not just a UX issue; it changes the security boundary because tired users may approve too much or disable prompts entirely.
• Task-scoped temporary permissions can preserve agent momentum while limiting standing access.
• Review prompts should distinguish safe observation, low-impact execution, credential use, customer-data access, money movement, and irreversible actions.
• Auditability matters after approval because users and organizations need to reconstruct what the agent did when permission was granted.
• Enterprise systems need to design for changing trust levels: an agent may earn more autonomy in a narrow workflow without earning broad authority over unrelated systems.

Connections

• Agent Permission Boundaries — broader authority and blast-radius frame.
• Enterprise Agent Governance — organization-level review and audit context.
• Agent Runtime Execution Layer — infrastructure layer that can enforce temporary scopes and logs.
• Agent Identity And Authentication — attribution layer for deciding which actor took an approved action.
• Agent Spend Controls / 智能体消费控制 — payment and budget form of approval scoping.
• Human Judgment Under AI — human responsibility boundary that approval prompts are supposed to preserve.