A team is designing a multi-agent customer-success system. They're debating: should they build one generalist agent that handles billing questions, technical support, and feature requests — or three specialist agents (Billing, Tech Support, Features) coordinated by a router?
The generalist can be implemented in a single prompt; the specialists require coordination overhead but each has tighter scope.
What's the org-chart-shaped framing?
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.
Division of labor in agent design follows the same logic as in organizations. When the domains genuinely require different expertise (different tools, different escalation contacts, different judgment standards), specialists pay off — the routing overhead is worth the tighter scope. When the domains share most of the substrate (same tools, same escalation, same judgment), specialization adds coordination cost without compensating benefit. The org-chart analogy isn't metaphor; it's the same optimization in a different medium.
Narrow agents do outperform generalists when the narrowness corresponds to genuine domain differences. When the domains overlap substantially, three "specialists" are really three near-copies of the same agent with different labels — paying coordination cost for no real specialization.
Coordination cost is a real consideration, but it's not the only one. A generalist agent with too broad a scope ends up vague, inconsistent, and over-routed-to-humans. The right comparison is total system effectiveness, not just operational simplicity.
This is the genuinely tempting wrong answer because it appeals to the company's existing optimization — the org chart isn't arbitrary; it reflects real choices about which domains needed separate ownership. But human org structure reflects human constraints: cognitive limits, accountability scope, career paths, span-of-control limits on managers. Agents don't share those constraints. The right specialization boundary for agents is determined by the architectural axes (different tools, different escalation paths, different judgment criteria), not by the staffing pattern that human limits produced.