Before a personal agent sees a founder's inbox, I want a one-page answer to what it may read, draft, and do.
Muse can prepare a weekly queue from connected work. The useful service is mapping each connected app to an owner, an approved action level, and a review path.
For a first pilot, start with one input, one output, and one clear stop.
The permission map has three columns: read, draft for approval, and human-only. A calendar move or message send does not silently graduate from draft to action.
Give the agent a request that looks routine but changes a commitment. It should stop and ask the named owner.
Which inbox or calendar action would you keep human-only?