Really cool approach, I like the “Unix philosophy” for agents. Curious how you handle state persistence and chaining sub-agents when agents are depth-limited. Also, do you have any strategies for ensuring data consistency across runs, especially when multiple agents interact with the same files?