One practical lesson with agents that control physical things: give them a small set of named actions with hard limits, such as 'water zone 2 for at most 10 minutes', instead of raw device access. Models sometimes misread a sensor value or repeat an action because they forgot they already did it. A log of past actions that the agent reads before every decision, plus a cap on how often each action can run per day, turns most of those mistakes into harmless no-ops. A plant can survive one bad guess but not a stuck valve.