A team scaled their customer success system from one generalist agent to eleven specialists (Billing, Tech Support, Features, Onboarding, Renewals, Churn, Account Health, Compliance, Localization, Enterprise, SMB). Coordination logic now consumes more development time than the agents themselves. Routing errors are common. End-to-end latency has tripled. Quality on individual queries is better than the generalist days, but throughput is worse.
What's the right diagnosis?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Coordination cost scales worse than linearly with agent count. Eleven specialists means a routing layer that has to disambiguate between eleven possibilities, handoff logic for each pair that needs handoffs, and an observability surface that spans eleven components. At some point the overhead outweighs the marginal benefit of further specialization. The org-chart analogy holds: most companies don't have eleven separate customer-success departments because the coordination cost would crush the specialization benefit. Consolidate to the level where specialization is actually load-bearing.
Individual agent quality is fine — that's what the scenario says. The problem is system-level: coordination, routing, latency. Optimizing individual agents doesn't address system-level cost.
A better orchestrator helps at the margin, but eleven-way routing is inherently more expensive than three-way routing — at some point the orchestrator itself becomes the bottleneck. The fix is reducing the count, not optimizing the routing.
Model load doesn't triple latency — three additional handoffs each adding seconds of round-trip latency does. The latency growth tracks the agent count, not the model.