How do you let non-engineers approve risky agent actions without killing the flow?
A builder asks how to let non-engineers approve risky agent actions: 'ask first' flags plus a human inbox, draft and live kept separate, and open questions on batching approvals, handling silence, and keeping audit logs useful. The thread is a practical look at human-in-the-loop design for agency agents, including the gap of shared OAuth tokens across an agent's users.