What's the right delegation model for human-agent work?

When should the human specify precisely and when should the agent decide? This is about finding the right handoff points that feel natural rather than bureaucratic.