Counter staff need to watch tokens stream while an agent drafts a denial letter so they can co-edit live. What development choice best supports that workflow?

Counter staff drafting sensitive municipal correspondence, such as denial letters, need human–agent co-authoring rather than an opaque finished product. Streaming the draft in real time into a staff UI that allows live co-editing before a resident sees a final letter is the best fit: tokens appear as they arrive, so you can catch tone, citation, and eligibility mistakes immediately, while co-edit controls keep staff accountable for the official artifact. Residents should receive only the staff-approved version, preserving progressive reveal inside the counter workflow without exposing unfinished model output. Hiding the draft until a supervisor batch-approves overnight printouts removes live oversight and can delay time-sensitive denials or appeals. Sending the finished letter by email after generation, with no live review or edits, removes the human gate on sensitive civic language. Requiring staff to retype the letter from a voice memo adds transcription error, wastes labor, and forgoes the interactive streaming workflow. Exam caveat: a staff-facing stream-and-edit process is not resident-facing auto-send; sensitive municipal letters typically need a human approval boundary even when streaming is enabled. Practical check: start generating a denial letter in the staff co-edit UI, interrupt and edit a clause mid-stream, confirm the resident channel remains unpublished until staff explicitly finalize, and verify the saved official letter matches the edited staff version rather than the raw unfinished stream.