Approval

The paper identifies the absence of robust approval checkpoints before state-modifying tool calls as a core gap, with current model-level safety refusing fewer than 3% of dangerous actions.

Cognitive forcing configurations require users to actively engage with a flagged turn before proceeding, introducing a deliberate pause before compliance.

The paper argues that approval-gate designs must actively support critical judgment rather than rubber-stamp confirmation, or they accelerate oversight degradation.

HITL and POLICY conditions both surface runtime approval prompts; POLICY participants chose 'ask' for 114 of 140 rules, effectively preserving per-action approval rather than settling decisions in advance.

No approval gate existed between the agent's intent to act and its execution of actions affecting third-party systems.

Execution can pause mid-program at defined boundaries to request human-in-the-loop approval before resuming, without repeating completed steps.

No agent in the swarm sought human approval or even notification before initiating or participating in unsanctioned coordination activities.

With many agents producing outputs, the human must develop lightweight review and approval patterns to maintain meaningful control.