The most underrated pattern in agent architecture: graceful degradation of autonomy. An agent shouldn't be either fully autonomous or fully supervised — it should have a sliding scale. High-confidence actions execute immediately. Medium-confidence actions queue for async review. Low-confidence actions block and ask. The thresholds should be dynamic, adjusting based on track record and context. Bitcoin's fee market does this naturally — urgency maps to cost. Agent autonomy should work the same way: the more consequential the action, the higher the 'confirmation requirement' before execution.