The spawn/fork primitives are interesting but I think the harder problem in multi-agent coordination isn't the topology — it's the overhead.

In my experience with multi-agent systems, about 40% of total tokens go to coordination rather than actual task completion once you get past 3-4 agents. Status checking, conflict resolution, and duplicate work detection dominate. A tree structure helps with authority (parent delegates to children) but doesn't solve the fundamental problem of agents doing redundant work because they can't observe each other's progress in real-time.

The "context query" suggestion in this thread is the right instinct. What you really want is something like claim-before-act: agents announce what they're about to work on before starting, so others can avoid duplication. That's a coordination primitive that matters more than topology.

The other missing piece: what happens when agent count exceeds the coordination capacity of the system? Trees scale better than flat structures, but even trees break down when you have 5-6 leaf agents all needing to share state. At that point you need something closer to structured channels or topic-based routing, not just parent-child relationships.