Orchestrate Multiple Agents to Resolve a Cross-System Incident

Consider a customer who has paid for a software subscription but has lost access to the workspace. The customer sees a downgraded or blocked account. The expected state is full workspace access. Resolving the issue requires information from several systems that do not automatically tell the same story.

The relevant sources include:

• Contract terms or another policy store defining what the customer is entitled to receive.
• Billing records containing invoices, payments, account status, plans, and seat information.
• Support records describing what the customer is experiencing and when the issue was reported.
• Workspace or entitlement records containing the current access level, flags, seat limits, and product state.
• A time-ordered event history showing when each state changed.

No single source is sufficient. The contract may say the customer should have access. The payment system may show that the invoice was paid. The billing account may still be marked past due. The workspace database may show a downgraded entitlement. The task is to reconstruct the sequence that produced the mismatch.

A multi-agent design assigns each investigation to a specialist.

The support agent begins with the customer-visible problem. It determines what the customer is seeing, where the issue occurs, and what was reported. It searches only the support-related memory available to it.

The billing agent investigates the transaction and account state. It checks the invoice, payment capture, reconciliation status, plan, seats, and the billing account. It determines whether the financial system believes the customer is paid, overdue, or in a grace period.

The workspace or entitlement agent checks the operational state. It determines what access is currently active, what changed, and whether the expected entitlement update was applied.

The supervisor agent receives the specialized findings and has access to the broader memory domains. Its role is to connect the evidence, compare the actual state with the intended state, and decide what action follows.

In the example, the combined evidence showed that the invoice was paid and the payment was captured, but the account remained past due because the reconciliation event did not clear the status correctly. Since the entitlement system depended on that status transition, the workspace access did not update. The corrective action was to resynchronize the entitlement state and restore full access. The preventive action was to improve event emission and reconciliation handling so that the systems remain aligned.

Contract terms act as a source of truth for what should be correct at each stage. Billing data explains the financial state. Entitlement data explains the actual product state. The timeline audit stream organizes the events chronologically so the supervisor can identify the point at which the systems diverged.

The supervisor does not replace the specialized agents. It integrates their findings. This separation allows each agent to operate with limited access while still enabling a final cross-domain explanation.

The same pattern applies beyond customer support. Any incident involving multiple systems can use specialized agents to investigate local evidence and a supervisor to reconstruct the whole. The memory layer is what allows their findings to persist, remain scoped, and be combined without requiring every agent to read and write everything.