approval-workflows

Outlines six human-in-the-loop patterns for AI agents ordered by intervention level: log-only, digest review, confidence thresholds, draft-then-send, approval gates on side effects, and two-person approval for high-risk actions.

The article explores the distinction between approval and completion states in agent workflows through a simulation test, highlighting the importance of idempotency and proper handling of retries to avoid duplicate actions.

The article discusses the need for re-checking permissions and actions immediately before an AI agent executes a task in production, due to potential changes in conditions like record states or approvals.

Discusses how human approval in AI agent workflows can be meaningless if the approved action differs from the executed one, urging stricter binding between approval and final action.

反思AI代理审批流程中人类把关的有效性，提出将小项审批批量处理为每日摘要以保持关注度，而非即时逐项审批。