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?
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?