its really interesting. Running agents close to the user’s data solves a lot of latency and privacy issues. The hard part seems to be balancing autonomy with predictability — once agents start chaining actions locally, debugging behavior becomes tricky. Curious how you're approaching that.