A message in your chat: what the agent wants to do, why, on what evidence, with approve or decline a tap away. Your decision is logged, and declines teach, the agent learns what you rejected and why.
Approvals arrive where you already are, so the loop is seconds, not a dashboard pilgrimage. The request shows its reasoning and sources, because asking for trust without evidence is not how this house works.
Treat declines as feedback with a sentence attached: 'no, this customer is mid-escalation' becomes memory, and the next request is smarter for it.