Specialists and an orchestrator: how to design an Agent Catalog that scales past ten KPIs

The fastest way to make AI useless inside an enterprise is to wrap it as a single assistant. One assistant has one set of guardrails, one prompt, one ambiguous scope. It does fine on the first demo question and breaks on the second one that crosses a module boundary.

The pattern that scales is the opposite: many narrow specialist agents, each responsible for a single KPI, coordinated by an orchestrator that knows which specialists to call for which question.

Why one generalist agent breaks

The generalist fails at the seams. Three specific seams break first.

• Cross-module reasoning. "Why is working capital deteriorating?" needs receivables, payables, inventory and forecast. The generalist either hallucinates one of them or asks the user to clarify, which defeats the point.
• Governance. Each module has different RBAC scopes, different audit requirements, different data-residency constraints. A single agent with the union of all permissions is the security incident waiting to happen.
• Versioning and ownership. When the definition of "DSO" changes, exactly one team should own the change. A generalist agent has no owner. A specialist DSO agent has one.

The fix is structural, not prompt-level. No prompt engineering rescues a model that has been given the wrong tools.

The specialist pattern - narrow, versioned, guard-railed

DivetIQ ships an Agent Catalog organised by domain:

• Financial - Cash Position, DSO/Collections, DPO/Payment, Revenue & Margin, Budget Variance, Reconciliation, Forecast, Treasury, Subscription Revenue.
• Operational - Inventory Optimization, Supplier Performance, Asset Health, Maintenance, Process Performance, Order Fulfillment, Field Service.
• Customer - Customer Health, Pipeline Health, Service Performance, Marketing Performance, Onboarding.
• People - Workforce Capacity, Retention Risk, Compliance Training, Payroll Validation, Wellbeing.
• ESG - Carbon Footprint, Energy Optimization, Supplier Sustainability, CSRD Reporting.

Each specialist has four properties that make it safe to deploy.

1. A narrow goal. "Watch DSO against threshold, attribute variance to customers, recommend collections actions." Not "be helpful."
2. An owner. A named team is responsible for the agent's definition, its tools, its guardrails.
3. A version. Changes are explicit, deployable, rollback-able. The agent in production today is not the agent that ran last quarter unless that is on purpose.
4. A kill switch. Globally, per agent, and per customer. EU AI Act–ready transparency notices included.

Specialists do not access databases directly. They use the same headless OpenAPI tools as human users, scoped by the same RBAC/ABAC. Every tool call is logged.

What the orchestrator actually does

The orchestrator is the meta-agent that handles questions no single specialist owns. "Why is working capital deteriorating?" routes to a plan: invoke Cash Position for the headline KPI, DSO for receivables variance, DPO for payables variance, Inventory Optimization for stock-tied capital. Each specialist returns its slice with lineage. The orchestrator recomposes the answer with attribution by source.

For complex playbooks - the working-capital recovery flow, the customer-retention playbook, the production yield response - the orchestrator drives a multi-step plan: detect, attribute, plan, optimize, approve, execute, learn. Each step is a specialist call. Each call is auditable. The human in the loop sits at the approve step by default.

V2 adds an Agent Marketplace: customers and partners can publish new specialists, vetted against the same guardrail standard. The same pattern, more catalog. The same architecture, more reach.

The discipline that makes the catalog scale is the discipline that makes any team scale: narrow scopes, clear ownership, explicit versions, a supervisor who only intervenes when the seams need crossing. Build the catalog that way and the eleventh KPI is no harder than the second.