When to reach for it
- The number of agents is large.
- Domains must be organized into sub-teams.
- Local decisions need to be aggregated centrally.
Hierarchical Supervisor.
Supervisors of supervisors. Scale by layers.
Multiple supervisors organize agents in hierarchical layers, delegating responsibility for larger teams or complex domains.
When it backfires
- A small number of agents suffices.
- Communication paths must remain short.
- Responsibilities cannot be cleanly separated.
The tradeoff
Scalable organizational structure is achieved against higher complexity and longer decision paths.
A shape you can draw on a napkin.
A top supervisor delegates to sub-supervisors, each managing its own team. The 'supervisor of supervisors' pattern.
Two ways this pattern will hurt you.
Decision latency multiplies with depth
A request goes through three supervisor layers before reaching a worker. Each layer adds a full LLM call.
Fix · Compress layers: only add a new supervisor tier when the fan-out exceeds a threshold (e.g., > 8 workers).
Lower supervisors optimize local metrics at global cost
A sub-supervisor maximizes its own team's throughput but the global objective requires cross-team coordination.
Fix · Expose the global objective to every supervisor tier, or add a periodic global rebalancing pass.
Where Hierarchical Supervisor is native.
Next pattern