Checkpoint policies for autonomous execution

Control objective

Checkpoint policies for autonomous execution put a human in the loop at the moments that matter, so an agent cannot take consequential or irreversible actions on its own. In a bank this means an agent may draft a fund transfer, account closure or large credit decision but must pause for explicit human approval before committing it, and it may only run so many steps unsupervised before checking back. To implement, classify actions by reversibility and impact, insert mandatory approval gates before any irreversible action, enforce a per-task-type budget that caps how many autonomous steps the agent may take before requiring human sign-off, and log every checkpoint, approval and denial as auditable evidence. No statistical metric is supplied; compliance is gate-based. The threshold is that 100% of irreversible actions are human-approved and autonomous steps stay within the defined budget per task type. Any irreversible action executed without approval, or an agent exceeding its step budget, is a breach that must block the action and trigger review - the system should fail safe by halting rather than proceeding unsupervised.