Use hierarchical sub-agent trees, not a flat star, to break the multi-agent context saturation ceiling

AppScale·medium signal

A star topology (one lead fanning to N children) saturates at roughly N ≈ W/m agents because every child's output funnels back into one window; a hierarchical tree where each aggregation node enforces b·m ≤ W locally scales to N = b^L agents across L depth levels. The practical rule for builders fanning out large agent swarms: insert intermediate aggregator nodes that compact before rolling up, because per-level compaction is a scaling requirement, not an optimization.