AT-301i · Module 2
Hierarchical Coordination
4 min read
Flat coordination — one orchestrator managing all agents — works up to approximately 10-12 agents. Beyond that, the orchestrator becomes the bottleneck. Every message routes through the same point. Every conflict resolution, every priority decision, every escalation lands on one entity.
Hierarchical coordination adds a layer: sub-team leads handle intra-team coordination, and the system coordinator handles only inter-team coordination and strategic decisions. The sub-team lead manages 4-5 agents directly — well within the effective span of control. The system coordinator manages 4-5 sub-team leads — also within effective range. The hierarchy adds one hop to cross-team communications but removes the single-point bottleneck.
Implementation: CLAWMANDER operates as system coordinator. Sub-team leads are implicit — the most senior agent in each functional group handles intra-team routing. CLOSER leads the Revenue sub-team. QUILL leads the Content sub-team. SCOPE leads the Intelligence sub-team. This is not a formal hierarchy — it is a coordination topology that distributes the orchestration workload. The sub-team leads still report to CLAWMANDER for cross-team coordination, but they handle routine intra-team work autonomously.
Do This
- Add coordination layers when the orchestrator processes more than 100 messages per day
- Keep span of control at 4-6 per coordinator — cognitive limits apply to AI too
- Designate sub-team leads from existing senior agents — do not add coordination-only agents
Avoid This
- Add layers preemptively — at 8 agents, flat coordination still outperforms hierarchy
- Create coordination-only agents that produce no direct output — they become bureaucracy
- Skip the hierarchy and add a second system coordinator — that creates a split-brain problem