How are you designing human review for production AI agents?
For people running agents in real workflows: what do you let the agent do automatically, what requires human approval, and how do you keep the review process auditable?
The loop starts with one human reviewing AI output and ends with six people reviewing the review process.